Spaces:
Running
Running
Commit ·
ebabb49
1
Parent(s): 6f608c8
v0.8.8 LongScore mode — anti-bullshit pack #14 + Hub badge readability fix
Browse files22nd mode: 🎯 LongScore. Lookup any HF model id → see relative degradation
past short context, sourced from RULER per-length + HELMET aggregate.
Why: every model claims a 128K context window, but accuracy degrades long
before that. Raw long-ctx scores are dominated by base ability — a smarter
model with a worse long-ctx recipe still scores higher than a less-smart
model with a better one, hiding the actual degradation. The 100-LongBench
paper (ACL 2025, arXiv:2505.19293) proposed LongScore to disentangle base
ability from true long-ctx capability:
LC_l = (S_l − Base) / Base
Base = mean(S_short)
LongScore = mean(LC_l for l in {16K, 32K, 64K, 128K})
0 = no degradation; -0.30 = severe.
Stage 0.5 verified (per PROTOCOL.md): no existing browser tool surfaces
LongScore for a given model id. tiktokenizer.vercel.app, llm-stats.com,
BenchLM.ai use raw RULER/needle scores. HELM Long Context aggregates 5
benchmarks but doesn't compute LongScore. Genuinely novel.
Files:
- data/longscore_kb.json (NEW, ~70 KB) — 93 unique models keyed by
canonical HF id. 35 with full LongScore (RULER per-length present).
Source: ruler_kb_day5.json + helmet_kb.json from large_model_validation/
Day-5 + Day-8 work. Built by build_longscore_kb_for_tafagent.py.
- js/longscore.js (NEW) — pure logic: loadKB, normalize, classify,
lookup, listAllIds, rank. No UI strings. ES module.
- index.html — new tab 🎯 LongScore between Token Tax and Solutions Hub.
New <section id="longscore-section"> with input + 3 example buttons +
output panels. Help modal v0.8.8 entry. Inventory v0.8.8 entry.
- js/main.js — initLongscore, renderLongscoreResult (per-length bars +
HELMET 7-task collapsible + verdict color band), runLongscoreLookup,
button wiring + Enter-key handler.
- js/i18n.js — 35 new keys × 4 langs (EN/ES/FR/ZH) = 140 keys total.
Includes mode label, tooltip, formula note, miss/hit/helmet_only state
text, verdict ladder (no_degradation → mild → moderate → severe →
extreme), help modal body.
- data/solutions_hub.json — new pain entry "long_ctx_degradation"
covered by 🎯 LongScore. Curated 6 external tools: 100-LongBench
paper, HELMET (repo + sheet), RULER, LongBench v2, Chroma context-rot.
- scripts/test_longscore.mjs (NEW) — 25 smoke tests for normalize,
classify, lookup. All pass.
- scripts/test_longscore_e2e.mjs (NEW) — 5 E2E lookup cases for the
3 example buttons + HELMET-only model + miss. All pass.
Hub badge fix (separate but bundled, 1-line):
- The "covered by mode" badges in Solutions Hub had `color: var(--success)`
inherited from .badge default and were also given inline
`background:#3fb950` (also success green). Result: green text on green
background → invisible. Fixed by adding inline `color:#fff` (white) and
`border-color` to all 3 badge variants (covered, planned, external).
Verification:
- node scripts/test_longscore.mjs → 25/25 pass
- node scripts/test_longscore_e2e.mjs → 5/5 pass
- python -m http.server + curl all assets → HTTP 200
- KB sanity: Llama-3.1-70B-Inst LongScore = -0.1024 matches Day-8
manual computation exactly.
Live HF Space verification PENDING (requires push to origin).
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
- data/longscore_kb.json +2941 -0
- data/solutions_hub.json +16 -0
- index.html +33 -0
- js/i18n.js +152 -0
- js/longscore.js +105 -0
- js/main.js +171 -4
- scripts/test_longscore.mjs +72 -0
- scripts/test_longscore_e2e.mjs +34 -0
data/longscore_kb.json
ADDED
|
@@ -0,0 +1,2941 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"version": "v0.8.8-longscore-2026-05-08",
|
| 3 |
+
"metric": "LongScore (100-LongBench, ACL 2025, arXiv:2505.19293, §3.2)",
|
| 4 |
+
"metric_formula": "Base = mean(S_4K, S_8K); LC_l = (S_l - Base) / Base; LongScore = mean(LC_l for l in {16K, 32K, 64K, 128K})",
|
| 5 |
+
"metric_interpretation": {
|
| 6 |
+
"0.0": "no degradation past short context",
|
| 7 |
+
"-0.05": "mild degradation (~5% relative drop)",
|
| 8 |
+
"-0.15": "moderate degradation (~15% relative drop)",
|
| 9 |
+
"-0.30": "severe degradation (~30% relative drop)",
|
| 10 |
+
"negative": "more negative = worse long-ctx retention"
|
| 11 |
+
},
|
| 12 |
+
"thresholds": {
|
| 13 |
+
"no_degradation": -0.02,
|
| 14 |
+
"mild": -0.1,
|
| 15 |
+
"moderate": -0.2,
|
| 16 |
+
"severe": -0.3
|
| 17 |
+
},
|
| 18 |
+
"sources": {
|
| 19 |
+
"ruler": "NVIDIA RULER leaderboard + Qwen2.5 Tech Report Table 16 (n=35)",
|
| 20 |
+
"helmet": "HELMET Google Sheet (princeton-nlp; arXiv:2410.02694; n=63)"
|
| 21 |
+
},
|
| 22 |
+
"stats": {
|
| 23 |
+
"n_total": 93,
|
| 24 |
+
"n_ruler_only": 30,
|
| 25 |
+
"n_helmet_only": 58,
|
| 26 |
+
"n_both": 5,
|
| 27 |
+
"n_with_longscore": 35
|
| 28 |
+
},
|
| 29 |
+
"models": {
|
| 30 |
+
"qwen-2-5-7-b-instruct": {
|
| 31 |
+
"display_name": "qwen2.5-7b-instruct",
|
| 32 |
+
"ruler_per_ctx": {
|
| 33 |
+
"4k": 96.7,
|
| 34 |
+
"8k": 95.1,
|
| 35 |
+
"16k": 93.7,
|
| 36 |
+
"32k": 89.4,
|
| 37 |
+
"64k": 74.5,
|
| 38 |
+
"128k": 31.4
|
| 39 |
+
},
|
| 40 |
+
"ruler_long_score": {
|
| 41 |
+
"base": 95.9,
|
| 42 |
+
"per_length_lc": {
|
| 43 |
+
"16k": -0.0229,
|
| 44 |
+
"32k": -0.0678,
|
| 45 |
+
"64k": -0.2231,
|
| 46 |
+
"128k": -0.6726
|
| 47 |
+
},
|
| 48 |
+
"avg_lc": -0.2466
|
| 49 |
+
},
|
| 50 |
+
"helmet": {
|
| 51 |
+
"overall": 22.8,
|
| 52 |
+
"categories": {
|
| 53 |
+
"Recall": 11.4,
|
| 54 |
+
"RAG": 30.6,
|
| 55 |
+
"Cite": 3.1,
|
| 56 |
+
"Re-rank": 1.6,
|
| 57 |
+
"ICL": 72.0,
|
| 58 |
+
"LongQA": 21.9,
|
| 59 |
+
"Summ": 18.8
|
| 60 |
+
},
|
| 61 |
+
"context_window_k": 128.0,
|
| 62 |
+
"params_b": "7.0",
|
| 63 |
+
"type": "♯"
|
| 64 |
+
},
|
| 65 |
+
"recipe_class": "native",
|
| 66 |
+
"params_b": 7,
|
| 67 |
+
"native_context_k": 32,
|
| 68 |
+
"source": "ruler+helmet"
|
| 69 |
+
},
|
| 70 |
+
"qwen-2-5-14-b-instruct": {
|
| 71 |
+
"display_name": "qwen2.5-14b-instruct",
|
| 72 |
+
"ruler_per_ctx": {
|
| 73 |
+
"4k": 96.9,
|
| 74 |
+
"8k": 97.1,
|
| 75 |
+
"16k": 95.5,
|
| 76 |
+
"32k": 95.5,
|
| 77 |
+
"64k": 90.3,
|
| 78 |
+
"128k": 82.0
|
| 79 |
+
},
|
| 80 |
+
"ruler_long_score": {
|
| 81 |
+
"base": 97.0,
|
| 82 |
+
"per_length_lc": {
|
| 83 |
+
"16k": -0.0155,
|
| 84 |
+
"32k": -0.0155,
|
| 85 |
+
"64k": -0.0691,
|
| 86 |
+
"128k": -0.1546
|
| 87 |
+
},
|
| 88 |
+
"avg_lc": -0.0637
|
| 89 |
+
},
|
| 90 |
+
"helmet": null,
|
| 91 |
+
"recipe_class": "native",
|
| 92 |
+
"params_b": 14,
|
| 93 |
+
"native_context_k": 32,
|
| 94 |
+
"source": "ruler"
|
| 95 |
+
},
|
| 96 |
+
"qwen-2-5-32-b-instruct": {
|
| 97 |
+
"display_name": "qwen2.5-32b-instruct",
|
| 98 |
+
"ruler_per_ctx": {
|
| 99 |
+
"4k": 97.7,
|
| 100 |
+
"8k": 97.2,
|
| 101 |
+
"16k": 97.7,
|
| 102 |
+
"32k": 96.5,
|
| 103 |
+
"64k": 88.5,
|
| 104 |
+
"128k": 67.0
|
| 105 |
+
},
|
| 106 |
+
"ruler_long_score": {
|
| 107 |
+
"base": 97.45,
|
| 108 |
+
"per_length_lc": {
|
| 109 |
+
"16k": 0.0026,
|
| 110 |
+
"32k": -0.0097,
|
| 111 |
+
"64k": -0.0918,
|
| 112 |
+
"128k": -0.3125
|
| 113 |
+
},
|
| 114 |
+
"avg_lc": -0.1028
|
| 115 |
+
},
|
| 116 |
+
"helmet": null,
|
| 117 |
+
"recipe_class": "native",
|
| 118 |
+
"params_b": 32,
|
| 119 |
+
"native_context_k": 32,
|
| 120 |
+
"source": "ruler"
|
| 121 |
+
},
|
| 122 |
+
"mistral-7-b-v-0-2": {
|
| 123 |
+
"display_name": "mistral-7b-v0.2",
|
| 124 |
+
"ruler_per_ctx": {
|
| 125 |
+
"4k": 93.6,
|
| 126 |
+
"8k": 91.2,
|
| 127 |
+
"16k": 87.2,
|
| 128 |
+
"32k": 75.4,
|
| 129 |
+
"64k": 49.0,
|
| 130 |
+
"128k": 13.8
|
| 131 |
+
},
|
| 132 |
+
"ruler_long_score": {
|
| 133 |
+
"base": 92.4,
|
| 134 |
+
"per_length_lc": {
|
| 135 |
+
"16k": -0.0563,
|
| 136 |
+
"32k": -0.184,
|
| 137 |
+
"64k": -0.4697,
|
| 138 |
+
"128k": -0.8506
|
| 139 |
+
},
|
| 140 |
+
"avg_lc": -0.3901
|
| 141 |
+
},
|
| 142 |
+
"helmet": null,
|
| 143 |
+
"recipe_class": "native",
|
| 144 |
+
"params_b": 7,
|
| 145 |
+
"native_context_k": 32,
|
| 146 |
+
"source": "ruler"
|
| 147 |
+
},
|
| 148 |
+
"mixtral-8-x-7-b": {
|
| 149 |
+
"display_name": "mixtral-8x7b",
|
| 150 |
+
"ruler_per_ctx": {
|
| 151 |
+
"4k": 94.9,
|
| 152 |
+
"8k": 92.1,
|
| 153 |
+
"16k": 92.5,
|
| 154 |
+
"32k": 85.9,
|
| 155 |
+
"64k": 72.4,
|
| 156 |
+
"128k": 44.5
|
| 157 |
+
},
|
| 158 |
+
"ruler_long_score": {
|
| 159 |
+
"base": 93.5,
|
| 160 |
+
"per_length_lc": {
|
| 161 |
+
"16k": -0.0107,
|
| 162 |
+
"32k": -0.0813,
|
| 163 |
+
"64k": -0.2257,
|
| 164 |
+
"128k": -0.5241
|
| 165 |
+
},
|
| 166 |
+
"avg_lc": -0.2104
|
| 167 |
+
},
|
| 168 |
+
"helmet": null,
|
| 169 |
+
"recipe_class": "native",
|
| 170 |
+
"params_b": 47,
|
| 171 |
+
"native_context_k": 32,
|
| 172 |
+
"source": "ruler"
|
| 173 |
+
},
|
| 174 |
+
"mixtral-8-x-22-b-instruct": {
|
| 175 |
+
"display_name": "mixtral-8x22b-instruct",
|
| 176 |
+
"ruler_per_ctx": {
|
| 177 |
+
"4k": 95.6,
|
| 178 |
+
"8k": 94.9,
|
| 179 |
+
"16k": 93.4,
|
| 180 |
+
"32k": 90.9,
|
| 181 |
+
"64k": 84.7,
|
| 182 |
+
"128k": 31.7
|
| 183 |
+
},
|
| 184 |
+
"ruler_long_score": {
|
| 185 |
+
"base": 95.25,
|
| 186 |
+
"per_length_lc": {
|
| 187 |
+
"16k": -0.0194,
|
| 188 |
+
"32k": -0.0457,
|
| 189 |
+
"64k": -0.1108,
|
| 190 |
+
"128k": -0.6672
|
| 191 |
+
},
|
| 192 |
+
"avg_lc": -0.2108
|
| 193 |
+
},
|
| 194 |
+
"helmet": null,
|
| 195 |
+
"recipe_class": "native",
|
| 196 |
+
"params_b": 141,
|
| 197 |
+
"native_context_k": 64,
|
| 198 |
+
"source": "ruler"
|
| 199 |
+
},
|
| 200 |
+
"qwen-2-72-b": {
|
| 201 |
+
"display_name": "qwen2-72b",
|
| 202 |
+
"ruler_per_ctx": {
|
| 203 |
+
"4k": 96.9,
|
| 204 |
+
"8k": 96.1,
|
| 205 |
+
"16k": 94.9,
|
| 206 |
+
"32k": 94.1,
|
| 207 |
+
"64k": 79.8,
|
| 208 |
+
"128k": 53.7
|
| 209 |
+
},
|
| 210 |
+
"ruler_long_score": {
|
| 211 |
+
"base": 96.5,
|
| 212 |
+
"per_length_lc": {
|
| 213 |
+
"16k": -0.0166,
|
| 214 |
+
"32k": -0.0249,
|
| 215 |
+
"64k": -0.1731,
|
| 216 |
+
"128k": -0.4435
|
| 217 |
+
},
|
| 218 |
+
"avg_lc": -0.1645
|
| 219 |
+
},
|
| 220 |
+
"helmet": null,
|
| 221 |
+
"recipe_class": "native",
|
| 222 |
+
"params_b": 72,
|
| 223 |
+
"native_context_k": 128,
|
| 224 |
+
"source": "ruler"
|
| 225 |
+
},
|
| 226 |
+
"dbrx": {
|
| 227 |
+
"display_name": "dbrx",
|
| 228 |
+
"ruler_per_ctx": {
|
| 229 |
+
"4k": 95.1,
|
| 230 |
+
"8k": 93.8,
|
| 231 |
+
"16k": 83.6,
|
| 232 |
+
"32k": 63.1,
|
| 233 |
+
"64k": 2.4,
|
| 234 |
+
"128k": 0.0
|
| 235 |
+
},
|
| 236 |
+
"ruler_long_score": {
|
| 237 |
+
"base": 94.45,
|
| 238 |
+
"per_length_lc": {
|
| 239 |
+
"16k": -0.1149,
|
| 240 |
+
"32k": -0.3319,
|
| 241 |
+
"64k": -0.9746,
|
| 242 |
+
"128k": -1.0
|
| 243 |
+
},
|
| 244 |
+
"avg_lc": -0.6054
|
| 245 |
+
},
|
| 246 |
+
"helmet": null,
|
| 247 |
+
"recipe_class": "native",
|
| 248 |
+
"params_b": 132,
|
| 249 |
+
"native_context_k": 32,
|
| 250 |
+
"source": "ruler"
|
| 251 |
+
},
|
| 252 |
+
"qwen-1-5-72-b": {
|
| 253 |
+
"display_name": "qwen1.5-72b",
|
| 254 |
+
"ruler_per_ctx": {
|
| 255 |
+
"4k": 94.9,
|
| 256 |
+
"8k": 93.8,
|
| 257 |
+
"16k": 78.0,
|
| 258 |
+
"32k": 67.8,
|
| 259 |
+
"64k": 0.0,
|
| 260 |
+
"128k": 0.0
|
| 261 |
+
},
|
| 262 |
+
"ruler_long_score": {
|
| 263 |
+
"base": 94.35,
|
| 264 |
+
"per_length_lc": {
|
| 265 |
+
"16k": -0.1733,
|
| 266 |
+
"32k": -0.2814,
|
| 267 |
+
"64k": -1.0,
|
| 268 |
+
"128k": -1.0
|
| 269 |
+
},
|
| 270 |
+
"avg_lc": -0.6137
|
| 271 |
+
},
|
| 272 |
+
"helmet": null,
|
| 273 |
+
"recipe_class": "native",
|
| 274 |
+
"params_b": 72,
|
| 275 |
+
"native_context_k": 32,
|
| 276 |
+
"source": "ruler"
|
| 277 |
+
},
|
| 278 |
+
"mistral-large-2407": {
|
| 279 |
+
"display_name": "mistral-large-2407",
|
| 280 |
+
"ruler_per_ctx": {
|
| 281 |
+
"4k": 96.2,
|
| 282 |
+
"8k": 96.1,
|
| 283 |
+
"16k": 95.1,
|
| 284 |
+
"32k": 93.0,
|
| 285 |
+
"64k": 78.8,
|
| 286 |
+
"128k": 23.7
|
| 287 |
+
},
|
| 288 |
+
"ruler_long_score": {
|
| 289 |
+
"base": 96.15,
|
| 290 |
+
"per_length_lc": {
|
| 291 |
+
"16k": -0.0109,
|
| 292 |
+
"32k": -0.0328,
|
| 293 |
+
"64k": -0.1804,
|
| 294 |
+
"128k": -0.7535
|
| 295 |
+
},
|
| 296 |
+
"avg_lc": -0.2444
|
| 297 |
+
},
|
| 298 |
+
"helmet": null,
|
| 299 |
+
"recipe_class": "native",
|
| 300 |
+
"params_b": 123,
|
| 301 |
+
"native_context_k": 32,
|
| 302 |
+
"source": "ruler"
|
| 303 |
+
},
|
| 304 |
+
"yi-34-b-200-k": {
|
| 305 |
+
"display_name": "yi-34b-200k",
|
| 306 |
+
"ruler_per_ctx": {
|
| 307 |
+
"4k": 93.3,
|
| 308 |
+
"8k": 92.2,
|
| 309 |
+
"16k": 91.3,
|
| 310 |
+
"32k": 87.5,
|
| 311 |
+
"64k": 83.2,
|
| 312 |
+
"128k": 77.3
|
| 313 |
+
},
|
| 314 |
+
"ruler_long_score": {
|
| 315 |
+
"base": 92.75,
|
| 316 |
+
"per_length_lc": {
|
| 317 |
+
"16k": -0.0156,
|
| 318 |
+
"32k": -0.0566,
|
| 319 |
+
"64k": -0.103,
|
| 320 |
+
"128k": -0.1666
|
| 321 |
+
},
|
| 322 |
+
"avg_lc": -0.0854
|
| 323 |
+
},
|
| 324 |
+
"helmet": {
|
| 325 |
+
"overall": 40.3,
|
| 326 |
+
"categories": {
|
| 327 |
+
"Recall": 74.1,
|
| 328 |
+
"RAG": 59.5,
|
| 329 |
+
"Cite": 1.7,
|
| 330 |
+
"Re-rank": 20.5,
|
| 331 |
+
"ICL": 85.0,
|
| 332 |
+
"LongQA": 26.6,
|
| 333 |
+
"Summ": 14.3
|
| 334 |
+
},
|
| 335 |
+
"context_window_k": 200.0,
|
| 336 |
+
"params_b": "34.0",
|
| 337 |
+
"type": "♭"
|
| 338 |
+
},
|
| 339 |
+
"recipe_class": "engineered_long_ctx",
|
| 340 |
+
"params_b": 34,
|
| 341 |
+
"native_context_k": 4,
|
| 342 |
+
"source": "ruler+helmet"
|
| 343 |
+
},
|
| 344 |
+
"command-r-35-b": {
|
| 345 |
+
"display_name": "command-r-35b",
|
| 346 |
+
"ruler_per_ctx": {
|
| 347 |
+
"4k": 93.8,
|
| 348 |
+
"8k": 93.3,
|
| 349 |
+
"16k": 92.4,
|
| 350 |
+
"32k": 89.5,
|
| 351 |
+
"64k": 84.9,
|
| 352 |
+
"128k": 76.0
|
| 353 |
+
},
|
| 354 |
+
"ruler_long_score": {
|
| 355 |
+
"base": 93.55,
|
| 356 |
+
"per_length_lc": {
|
| 357 |
+
"16k": -0.0123,
|
| 358 |
+
"32k": -0.0433,
|
| 359 |
+
"64k": -0.0925,
|
| 360 |
+
"128k": -0.1876
|
| 361 |
+
},
|
| 362 |
+
"avg_lc": -0.0839
|
| 363 |
+
},
|
| 364 |
+
"helmet": null,
|
| 365 |
+
"recipe_class": "engineered_long_ctx",
|
| 366 |
+
"params_b": 35,
|
| 367 |
+
"native_context_k": 8,
|
| 368 |
+
"source": "ruler"
|
| 369 |
+
},
|
| 370 |
+
"command-r-0824-32-b": {
|
| 371 |
+
"display_name": "command-r-0824-32b",
|
| 372 |
+
"ruler_per_ctx": {
|
| 373 |
+
"4k": 94.7,
|
| 374 |
+
"8k": 93.7,
|
| 375 |
+
"16k": 93.1,
|
| 376 |
+
"32k": 90.8,
|
| 377 |
+
"64k": 86.6,
|
| 378 |
+
"128k": 74.7
|
| 379 |
+
},
|
| 380 |
+
"ruler_long_score": {
|
| 381 |
+
"base": 94.2,
|
| 382 |
+
"per_length_lc": {
|
| 383 |
+
"16k": -0.0117,
|
| 384 |
+
"32k": -0.0361,
|
| 385 |
+
"64k": -0.0807,
|
| 386 |
+
"128k": -0.207
|
| 387 |
+
},
|
| 388 |
+
"avg_lc": -0.0839
|
| 389 |
+
},
|
| 390 |
+
"helmet": null,
|
| 391 |
+
"recipe_class": "engineered_long_ctx",
|
| 392 |
+
"params_b": 32,
|
| 393 |
+
"native_context_k": 8,
|
| 394 |
+
"source": "ruler"
|
| 395 |
+
},
|
| 396 |
+
"command-r-plus-104-b": {
|
| 397 |
+
"display_name": "command-r-plus-104b",
|
| 398 |
+
"ruler_per_ctx": {
|
| 399 |
+
"4k": 95.6,
|
| 400 |
+
"8k": 95.2,
|
| 401 |
+
"16k": 94.2,
|
| 402 |
+
"32k": 92.0,
|
| 403 |
+
"64k": 84.3,
|
| 404 |
+
"128k": 63.1
|
| 405 |
+
},
|
| 406 |
+
"ruler_long_score": {
|
| 407 |
+
"base": 95.4,
|
| 408 |
+
"per_length_lc": {
|
| 409 |
+
"16k": -0.0126,
|
| 410 |
+
"32k": -0.0356,
|
| 411 |
+
"64k": -0.1164,
|
| 412 |
+
"128k": -0.3386
|
| 413 |
+
},
|
| 414 |
+
"avg_lc": -0.1258
|
| 415 |
+
},
|
| 416 |
+
"helmet": null,
|
| 417 |
+
"recipe_class": "engineered_long_ctx",
|
| 418 |
+
"params_b": 104,
|
| 419 |
+
"native_context_k": 8,
|
| 420 |
+
"source": "ruler"
|
| 421 |
+
},
|
| 422 |
+
"command-r-plus-0824-104-b": {
|
| 423 |
+
"display_name": "command-r-plus-0824-104b",
|
| 424 |
+
"ruler_per_ctx": {
|
| 425 |
+
"4k": 96.0,
|
| 426 |
+
"8k": 95.1,
|
| 427 |
+
"16k": 94.0,
|
| 428 |
+
"32k": 92.4,
|
| 429 |
+
"64k": 85.4,
|
| 430 |
+
"128k": 64.6
|
| 431 |
+
},
|
| 432 |
+
"ruler_long_score": {
|
| 433 |
+
"base": 95.55,
|
| 434 |
+
"per_length_lc": {
|
| 435 |
+
"16k": -0.0162,
|
| 436 |
+
"32k": -0.033,
|
| 437 |
+
"64k": -0.1062,
|
| 438 |
+
"128k": -0.3239
|
| 439 |
+
},
|
| 440 |
+
"avg_lc": -0.1198
|
| 441 |
+
},
|
| 442 |
+
"helmet": null,
|
| 443 |
+
"recipe_class": "engineered_long_ctx",
|
| 444 |
+
"params_b": 104,
|
| 445 |
+
"native_context_k": 8,
|
| 446 |
+
"source": "ruler"
|
| 447 |
+
},
|
| 448 |
+
"mistral-large-2411-123-b": {
|
| 449 |
+
"display_name": "mistral-large-2411-123b",
|
| 450 |
+
"ruler_per_ctx": {
|
| 451 |
+
"4k": 96.4,
|
| 452 |
+
"8k": 96.3,
|
| 453 |
+
"16k": 95.3,
|
| 454 |
+
"32k": 94.0,
|
| 455 |
+
"64k": 85.9,
|
| 456 |
+
"128k": 48.1
|
| 457 |
+
},
|
| 458 |
+
"ruler_long_score": {
|
| 459 |
+
"base": 96.35,
|
| 460 |
+
"per_length_lc": {
|
| 461 |
+
"16k": -0.0109,
|
| 462 |
+
"32k": -0.0244,
|
| 463 |
+
"64k": -0.1085,
|
| 464 |
+
"128k": -0.5008
|
| 465 |
+
},
|
| 466 |
+
"avg_lc": -0.1612
|
| 467 |
+
},
|
| 468 |
+
"helmet": null,
|
| 469 |
+
"recipe_class": "engineered_long_ctx",
|
| 470 |
+
"params_b": 123,
|
| 471 |
+
"native_context_k": 8,
|
| 472 |
+
"source": "ruler"
|
| 473 |
+
},
|
| 474 |
+
"chatglm-3-6-b-128-k": {
|
| 475 |
+
"display_name": "chatglm3-6b-128k",
|
| 476 |
+
"ruler_per_ctx": {
|
| 477 |
+
"4k": 87.8,
|
| 478 |
+
"8k": 83.4,
|
| 479 |
+
"16k": 78.6,
|
| 480 |
+
"32k": 69.9,
|
| 481 |
+
"64k": 56.0,
|
| 482 |
+
"128k": 42.0
|
| 483 |
+
},
|
| 484 |
+
"ruler_long_score": {
|
| 485 |
+
"base": 85.6,
|
| 486 |
+
"per_length_lc": {
|
| 487 |
+
"16k": -0.0818,
|
| 488 |
+
"32k": -0.1834,
|
| 489 |
+
"64k": -0.3458,
|
| 490 |
+
"128k": -0.5093
|
| 491 |
+
},
|
| 492 |
+
"avg_lc": -0.2801
|
| 493 |
+
},
|
| 494 |
+
"helmet": null,
|
| 495 |
+
"recipe_class": "aggressive_ext_small_base",
|
| 496 |
+
"params_b": 6,
|
| 497 |
+
"native_context_k": 2,
|
| 498 |
+
"source": "ruler"
|
| 499 |
+
},
|
| 500 |
+
"lwm-7-b": {
|
| 501 |
+
"display_name": "lwm-7b",
|
| 502 |
+
"ruler_per_ctx": {
|
| 503 |
+
"4k": 82.3,
|
| 504 |
+
"8k": 78.4,
|
| 505 |
+
"16k": 73.7,
|
| 506 |
+
"32k": 69.1,
|
| 507 |
+
"64k": 68.1,
|
| 508 |
+
"128k": 65.0
|
| 509 |
+
},
|
| 510 |
+
"ruler_long_score": {
|
| 511 |
+
"base": 80.35,
|
| 512 |
+
"per_length_lc": {
|
| 513 |
+
"16k": -0.0828,
|
| 514 |
+
"32k": -0.14,
|
| 515 |
+
"64k": -0.1525,
|
| 516 |
+
"128k": -0.191
|
| 517 |
+
},
|
| 518 |
+
"avg_lc": -0.1416
|
| 519 |
+
},
|
| 520 |
+
"helmet": null,
|
| 521 |
+
"recipe_class": "aggressive_ext_small_base",
|
| 522 |
+
"params_b": 7,
|
| 523 |
+
"native_context_k": 4,
|
| 524 |
+
"source": "ruler"
|
| 525 |
+
},
|
| 526 |
+
"glm-4-9-b-1-m": {
|
| 527 |
+
"display_name": "glm4-9b-1m",
|
| 528 |
+
"ruler_per_ctx": {
|
| 529 |
+
"4k": 94.7,
|
| 530 |
+
"8k": 92.8,
|
| 531 |
+
"16k": 92.1,
|
| 532 |
+
"32k": 89.9,
|
| 533 |
+
"64k": 86.7,
|
| 534 |
+
"128k": 83.1
|
| 535 |
+
},
|
| 536 |
+
"ruler_long_score": {
|
| 537 |
+
"base": 93.75,
|
| 538 |
+
"per_length_lc": {
|
| 539 |
+
"16k": -0.0176,
|
| 540 |
+
"32k": -0.0411,
|
| 541 |
+
"64k": -0.0752,
|
| 542 |
+
"128k": -0.1136
|
| 543 |
+
},
|
| 544 |
+
"avg_lc": -0.0619
|
| 545 |
+
},
|
| 546 |
+
"helmet": null,
|
| 547 |
+
"recipe_class": "aggressive_ext_small_base",
|
| 548 |
+
"params_b": 9,
|
| 549 |
+
"native_context_k": 8,
|
| 550 |
+
"source": "ruler"
|
| 551 |
+
},
|
| 552 |
+
"prolong-8-b-512-k": {
|
| 553 |
+
"display_name": "prolong-8b-512k",
|
| 554 |
+
"ruler_per_ctx": {
|
| 555 |
+
"4k": 94.5,
|
| 556 |
+
"8k": 92.5,
|
| 557 |
+
"16k": 92.3,
|
| 558 |
+
"32k": 89.3,
|
| 559 |
+
"64k": 83.2,
|
| 560 |
+
"128k": 81.6
|
| 561 |
+
},
|
| 562 |
+
"ruler_long_score": {
|
| 563 |
+
"base": 93.5,
|
| 564 |
+
"per_length_lc": {
|
| 565 |
+
"16k": -0.0128,
|
| 566 |
+
"32k": -0.0449,
|
| 567 |
+
"64k": -0.1102,
|
| 568 |
+
"128k": -0.1273
|
| 569 |
+
},
|
| 570 |
+
"avg_lc": -0.0738
|
| 571 |
+
},
|
| 572 |
+
"helmet": null,
|
| 573 |
+
"recipe_class": "aggressive_ext_small_base",
|
| 574 |
+
"params_b": 8,
|
| 575 |
+
"native_context_k": 8,
|
| 576 |
+
"source": "ruler"
|
| 577 |
+
},
|
| 578 |
+
"megabeam-mistral-7-b-512-k": {
|
| 579 |
+
"display_name": "megabeam-mistral-7b-512k",
|
| 580 |
+
"ruler_per_ctx": {
|
| 581 |
+
"4k": 93.8,
|
| 582 |
+
"8k": 92.5,
|
| 583 |
+
"16k": 92.0,
|
| 584 |
+
"32k": 89.2,
|
| 585 |
+
"64k": 83.7,
|
| 586 |
+
"128k": 83.7
|
| 587 |
+
},
|
| 588 |
+
"ruler_long_score": {
|
| 589 |
+
"base": 93.15,
|
| 590 |
+
"per_length_lc": {
|
| 591 |
+
"16k": -0.0123,
|
| 592 |
+
"32k": -0.0424,
|
| 593 |
+
"64k": -0.1014,
|
| 594 |
+
"128k": -0.1014
|
| 595 |
+
},
|
| 596 |
+
"avg_lc": -0.0644
|
| 597 |
+
},
|
| 598 |
+
"helmet": null,
|
| 599 |
+
"recipe_class": "aggressive_ext_small_base",
|
| 600 |
+
"params_b": 7,
|
| 601 |
+
"native_context_k": 32,
|
| 602 |
+
"source": "ruler"
|
| 603 |
+
},
|
| 604 |
+
"internlm-2-5-7-b-1-m": {
|
| 605 |
+
"display_name": "internlm2.5-7b-1m",
|
| 606 |
+
"ruler_per_ctx": {
|
| 607 |
+
"4k": 88.1,
|
| 608 |
+
"8k": 85.5,
|
| 609 |
+
"16k": 84.5,
|
| 610 |
+
"32k": 82.7,
|
| 611 |
+
"64k": 75.5,
|
| 612 |
+
"128k": 68.9
|
| 613 |
+
},
|
| 614 |
+
"ruler_long_score": {
|
| 615 |
+
"base": 86.8,
|
| 616 |
+
"per_length_lc": {
|
| 617 |
+
"16k": -0.0265,
|
| 618 |
+
"32k": -0.0472,
|
| 619 |
+
"64k": -0.1302,
|
| 620 |
+
"128k": -0.2062
|
| 621 |
+
},
|
| 622 |
+
"avg_lc": -0.1025
|
| 623 |
+
},
|
| 624 |
+
"helmet": null,
|
| 625 |
+
"recipe_class": "aggressive_ext_small_base",
|
| 626 |
+
"params_b": 7,
|
| 627 |
+
"native_context_k": 4,
|
| 628 |
+
"source": "ruler"
|
| 629 |
+
},
|
| 630 |
+
"gradientai-llama-3-70-b-1-m": {
|
| 631 |
+
"display_name": "gradientai-llama3-70b-1m",
|
| 632 |
+
"ruler_per_ctx": {
|
| 633 |
+
"4k": 95.1,
|
| 634 |
+
"8k": 94.4,
|
| 635 |
+
"16k": 90.8,
|
| 636 |
+
"32k": 85.4,
|
| 637 |
+
"64k": 80.9,
|
| 638 |
+
"128k": 72.1
|
| 639 |
+
},
|
| 640 |
+
"ruler_long_score": {
|
| 641 |
+
"base": 94.75,
|
| 642 |
+
"per_length_lc": {
|
| 643 |
+
"16k": -0.0417,
|
| 644 |
+
"32k": -0.0987,
|
| 645 |
+
"64k": -0.1462,
|
| 646 |
+
"128k": -0.2391
|
| 647 |
+
},
|
| 648 |
+
"avg_lc": -0.1314
|
| 649 |
+
},
|
| 650 |
+
"helmet": null,
|
| 651 |
+
"recipe_class": "aggressive_ext_small_base",
|
| 652 |
+
"params_b": 70,
|
| 653 |
+
"native_context_k": 8,
|
| 654 |
+
"source": "ruler"
|
| 655 |
+
},
|
| 656 |
+
"gradientai-llama-3-8-b-1-m": {
|
| 657 |
+
"display_name": "gradientai-llama3-8b-1m",
|
| 658 |
+
"ruler_per_ctx": {
|
| 659 |
+
"4k": 92.8,
|
| 660 |
+
"8k": 90.3,
|
| 661 |
+
"16k": 85.7,
|
| 662 |
+
"32k": 79.9,
|
| 663 |
+
"64k": 76.3,
|
| 664 |
+
"128k": 69.5
|
| 665 |
+
},
|
| 666 |
+
"ruler_long_score": {
|
| 667 |
+
"base": 91.55,
|
| 668 |
+
"per_length_lc": {
|
| 669 |
+
"16k": -0.0639,
|
| 670 |
+
"32k": -0.1273,
|
| 671 |
+
"64k": -0.1666,
|
| 672 |
+
"128k": -0.2409
|
| 673 |
+
},
|
| 674 |
+
"avg_lc": -0.1497
|
| 675 |
+
},
|
| 676 |
+
"helmet": null,
|
| 677 |
+
"recipe_class": "aggressive_ext_small_base",
|
| 678 |
+
"params_b": 8,
|
| 679 |
+
"native_context_k": 8,
|
| 680 |
+
"source": "ruler"
|
| 681 |
+
},
|
| 682 |
+
"longchat-7-b-32-k": {
|
| 683 |
+
"display_name": "longchat-7b-32k",
|
| 684 |
+
"ruler_per_ctx": {
|
| 685 |
+
"4k": 84.7,
|
| 686 |
+
"8k": 79.9,
|
| 687 |
+
"16k": 70.8,
|
| 688 |
+
"32k": 59.3,
|
| 689 |
+
"64k": 0.0,
|
| 690 |
+
"128k": 0.0
|
| 691 |
+
},
|
| 692 |
+
"ruler_long_score": {
|
| 693 |
+
"base": 82.3,
|
| 694 |
+
"per_length_lc": {
|
| 695 |
+
"16k": -0.1397,
|
| 696 |
+
"32k": -0.2795,
|
| 697 |
+
"64k": -1.0,
|
| 698 |
+
"128k": -1.0
|
| 699 |
+
},
|
| 700 |
+
"avg_lc": -0.6048
|
| 701 |
+
},
|
| 702 |
+
"helmet": null,
|
| 703 |
+
"recipe_class": "aggressive_ext_small_base",
|
| 704 |
+
"params_b": 7,
|
| 705 |
+
"native_context_k": 4,
|
| 706 |
+
"source": "ruler"
|
| 707 |
+
},
|
| 708 |
+
"longalpaca-13-b-32-k": {
|
| 709 |
+
"display_name": "longalpaca-13b-32k",
|
| 710 |
+
"ruler_per_ctx": {
|
| 711 |
+
"4k": 60.6,
|
| 712 |
+
"8k": 57.0,
|
| 713 |
+
"16k": 56.6,
|
| 714 |
+
"32k": 43.6,
|
| 715 |
+
"64k": 0.0,
|
| 716 |
+
"128k": 0.0
|
| 717 |
+
},
|
| 718 |
+
"ruler_long_score": {
|
| 719 |
+
"base": 58.8,
|
| 720 |
+
"per_length_lc": {
|
| 721 |
+
"16k": -0.0374,
|
| 722 |
+
"32k": -0.2585,
|
| 723 |
+
"64k": -1.0,
|
| 724 |
+
"128k": -1.0
|
| 725 |
+
},
|
| 726 |
+
"avg_lc": -0.574
|
| 727 |
+
},
|
| 728 |
+
"helmet": null,
|
| 729 |
+
"recipe_class": "aggressive_ext_small_base",
|
| 730 |
+
"params_b": 13,
|
| 731 |
+
"native_context_k": 4,
|
| 732 |
+
"source": "ruler"
|
| 733 |
+
},
|
| 734 |
+
"llama-3-1-70-b-instruct": {
|
| 735 |
+
"display_name": "llama3.1-70b-instruct",
|
| 736 |
+
"ruler_per_ctx": {
|
| 737 |
+
"4k": 96.5,
|
| 738 |
+
"8k": 95.8,
|
| 739 |
+
"16k": 95.4,
|
| 740 |
+
"32k": 94.8,
|
| 741 |
+
"64k": 88.4,
|
| 742 |
+
"128k": 66.6
|
| 743 |
+
},
|
| 744 |
+
"ruler_long_score": {
|
| 745 |
+
"base": 96.15,
|
| 746 |
+
"per_length_lc": {
|
| 747 |
+
"16k": -0.0078,
|
| 748 |
+
"32k": -0.014,
|
| 749 |
+
"64k": -0.0806,
|
| 750 |
+
"128k": -0.3073
|
| 751 |
+
},
|
| 752 |
+
"avg_lc": -0.1024
|
| 753 |
+
},
|
| 754 |
+
"helmet": {
|
| 755 |
+
"overall": 49.7,
|
| 756 |
+
"categories": {
|
| 757 |
+
"Recall": 90.7,
|
| 758 |
+
"RAG": 56.2,
|
| 759 |
+
"Cite": 7.5,
|
| 760 |
+
"Re-rank": 24.5,
|
| 761 |
+
"ICL": 81.4,
|
| 762 |
+
"LongQA": 56.3,
|
| 763 |
+
"Summ": 31.6
|
| 764 |
+
},
|
| 765 |
+
"context_window_k": 128.0,
|
| 766 |
+
"params_b": "70.0",
|
| 767 |
+
"type": "♯"
|
| 768 |
+
},
|
| 769 |
+
"recipe_class": "continued_pretrain_cliff",
|
| 770 |
+
"params_b": 70,
|
| 771 |
+
"native_context_k": 8,
|
| 772 |
+
"source": "ruler+helmet"
|
| 773 |
+
},
|
| 774 |
+
"llama-3-1-8-b-instruct": {
|
| 775 |
+
"display_name": "llama3.1-8b-instruct",
|
| 776 |
+
"ruler_per_ctx": {
|
| 777 |
+
"4k": 95.5,
|
| 778 |
+
"8k": 93.8,
|
| 779 |
+
"16k": 91.6,
|
| 780 |
+
"32k": 87.4,
|
| 781 |
+
"64k": 84.7,
|
| 782 |
+
"128k": 77.0
|
| 783 |
+
},
|
| 784 |
+
"ruler_long_score": {
|
| 785 |
+
"base": 94.65,
|
| 786 |
+
"per_length_lc": {
|
| 787 |
+
"16k": -0.0322,
|
| 788 |
+
"32k": -0.0766,
|
| 789 |
+
"64k": -0.1051,
|
| 790 |
+
"128k": -0.1865
|
| 791 |
+
},
|
| 792 |
+
"avg_lc": -0.1001
|
| 793 |
+
},
|
| 794 |
+
"helmet": {
|
| 795 |
+
"overall": 46.5,
|
| 796 |
+
"categories": {
|
| 797 |
+
"Recall": 95.2,
|
| 798 |
+
"RAG": 59.5,
|
| 799 |
+
"Cite": 2.9,
|
| 800 |
+
"Re-rank": 14.0,
|
| 801 |
+
"ICL": 83.9,
|
| 802 |
+
"LongQA": 43.2,
|
| 803 |
+
"Summ": 27.0
|
| 804 |
+
},
|
| 805 |
+
"context_window_k": 128.0,
|
| 806 |
+
"params_b": "8.0",
|
| 807 |
+
"type": "♯"
|
| 808 |
+
},
|
| 809 |
+
"recipe_class": "continued_pretrain_cliff",
|
| 810 |
+
"params_b": 8,
|
| 811 |
+
"native_context_k": 8,
|
| 812 |
+
"source": "ruler+helmet"
|
| 813 |
+
},
|
| 814 |
+
"mistral-nemo-12-b": {
|
| 815 |
+
"display_name": "mistral-nemo-12b",
|
| 816 |
+
"ruler_per_ctx": {
|
| 817 |
+
"4k": 87.8,
|
| 818 |
+
"8k": 87.2,
|
| 819 |
+
"16k": 87.7,
|
| 820 |
+
"32k": 69.0,
|
| 821 |
+
"64k": 46.8,
|
| 822 |
+
"128k": 19.0
|
| 823 |
+
},
|
| 824 |
+
"ruler_long_score": {
|
| 825 |
+
"base": 87.5,
|
| 826 |
+
"per_length_lc": {
|
| 827 |
+
"16k": 0.0023,
|
| 828 |
+
"32k": -0.2114,
|
| 829 |
+
"64k": -0.4651,
|
| 830 |
+
"128k": -0.7829
|
| 831 |
+
},
|
| 832 |
+
"avg_lc": -0.3643
|
| 833 |
+
},
|
| 834 |
+
"helmet": null,
|
| 835 |
+
"recipe_class": "continued_pretrain_cliff",
|
| 836 |
+
"params_b": 12,
|
| 837 |
+
"native_context_k": 8,
|
| 838 |
+
"source": "ruler"
|
| 839 |
+
},
|
| 840 |
+
"qwen-2-5-7-b-instruct-1-m": {
|
| 841 |
+
"display_name": "qwen2.5-7b-instruct-1m",
|
| 842 |
+
"ruler_per_ctx": {
|
| 843 |
+
"4k": 96.8,
|
| 844 |
+
"8k": 95.3,
|
| 845 |
+
"16k": 93.0,
|
| 846 |
+
"32k": 91.1,
|
| 847 |
+
"64k": 90.4,
|
| 848 |
+
"128k": 84.4
|
| 849 |
+
},
|
| 850 |
+
"ruler_long_score": {
|
| 851 |
+
"base": 96.05,
|
| 852 |
+
"per_length_lc": {
|
| 853 |
+
"16k": -0.0318,
|
| 854 |
+
"32k": -0.0515,
|
| 855 |
+
"64k": -0.0588,
|
| 856 |
+
"128k": -0.1213
|
| 857 |
+
},
|
| 858 |
+
"avg_lc": -0.0659
|
| 859 |
+
},
|
| 860 |
+
"helmet": null,
|
| 861 |
+
"recipe_class": "dca_ssm_flat",
|
| 862 |
+
"params_b": 7,
|
| 863 |
+
"native_context_k": 32,
|
| 864 |
+
"source": "ruler"
|
| 865 |
+
},
|
| 866 |
+
"qwen-2-5-14-b-instruct-1-m": {
|
| 867 |
+
"display_name": "qwen2.5-14b-instruct-1m",
|
| 868 |
+
"ruler_per_ctx": {
|
| 869 |
+
"4k": 97.5,
|
| 870 |
+
"8k": 97.1,
|
| 871 |
+
"16k": 94.6,
|
| 872 |
+
"32k": 94.9,
|
| 873 |
+
"64k": 94.9,
|
| 874 |
+
"128k": 92.2
|
| 875 |
+
},
|
| 876 |
+
"ruler_long_score": {
|
| 877 |
+
"base": 97.3,
|
| 878 |
+
"per_length_lc": {
|
| 879 |
+
"16k": -0.0277,
|
| 880 |
+
"32k": -0.0247,
|
| 881 |
+
"64k": -0.0247,
|
| 882 |
+
"128k": -0.0524
|
| 883 |
+
},
|
| 884 |
+
"avg_lc": -0.0324
|
| 885 |
+
},
|
| 886 |
+
"helmet": null,
|
| 887 |
+
"recipe_class": "dca_ssm_flat",
|
| 888 |
+
"params_b": 14,
|
| 889 |
+
"native_context_k": 32,
|
| 890 |
+
"source": "ruler"
|
| 891 |
+
},
|
| 892 |
+
"jamba-1-5-large": {
|
| 893 |
+
"display_name": "jamba-1.5-large",
|
| 894 |
+
"ruler_per_ctx": {
|
| 895 |
+
"4k": 96.7,
|
| 896 |
+
"8k": 96.6,
|
| 897 |
+
"16k": 96.4,
|
| 898 |
+
"32k": 96.0,
|
| 899 |
+
"64k": 95.4,
|
| 900 |
+
"128k": 95.1
|
| 901 |
+
},
|
| 902 |
+
"ruler_long_score": {
|
| 903 |
+
"base": 96.65,
|
| 904 |
+
"per_length_lc": {
|
| 905 |
+
"16k": -0.0026,
|
| 906 |
+
"32k": -0.0067,
|
| 907 |
+
"64k": -0.0129,
|
| 908 |
+
"128k": -0.016
|
| 909 |
+
},
|
| 910 |
+
"avg_lc": -0.0095
|
| 911 |
+
},
|
| 912 |
+
"helmet": null,
|
| 913 |
+
"recipe_class": "dca_ssm_flat",
|
| 914 |
+
"params_b": 398,
|
| 915 |
+
"native_context_k": 256,
|
| 916 |
+
"source": "ruler"
|
| 917 |
+
},
|
| 918 |
+
"jamba-1-5-mini": {
|
| 919 |
+
"display_name": "jamba-1.5-mini",
|
| 920 |
+
"ruler_per_ctx": {
|
| 921 |
+
"4k": 95.6,
|
| 922 |
+
"8k": 95.6,
|
| 923 |
+
"16k": 94.8,
|
| 924 |
+
"32k": 94.6,
|
| 925 |
+
"64k": 92.8,
|
| 926 |
+
"128k": 90.0
|
| 927 |
+
},
|
| 928 |
+
"ruler_long_score": {
|
| 929 |
+
"base": 95.6,
|
| 930 |
+
"per_length_lc": {
|
| 931 |
+
"16k": -0.0084,
|
| 932 |
+
"32k": -0.0105,
|
| 933 |
+
"64k": -0.0293,
|
| 934 |
+
"128k": -0.0586
|
| 935 |
+
},
|
| 936 |
+
"avg_lc": -0.0267
|
| 937 |
+
},
|
| 938 |
+
"helmet": {
|
| 939 |
+
"overall": 46.9,
|
| 940 |
+
"categories": {
|
| 941 |
+
"Recall": 90.0,
|
| 942 |
+
"RAG": 57.3,
|
| 943 |
+
"Cite": 3.1,
|
| 944 |
+
"Re-rank": 14.6,
|
| 945 |
+
"ICL": 91.0,
|
| 946 |
+
"LongQA": 54.2,
|
| 947 |
+
"Summ": 18.1
|
| 948 |
+
},
|
| 949 |
+
"context_window_k": 256.0,
|
| 950 |
+
"params_b": "12.0",
|
| 951 |
+
"type": "♯"
|
| 952 |
+
},
|
| 953 |
+
"recipe_class": "dca_ssm_flat",
|
| 954 |
+
"params_b": 52,
|
| 955 |
+
"native_context_k": 256,
|
| 956 |
+
"source": "ruler+helmet"
|
| 957 |
+
},
|
| 958 |
+
"phi-3-mini-3-8-b": {
|
| 959 |
+
"display_name": "phi3-mini-3.8b",
|
| 960 |
+
"ruler_per_ctx": {
|
| 961 |
+
"4k": 92.2,
|
| 962 |
+
"8k": 91.5,
|
| 963 |
+
"16k": 90.7,
|
| 964 |
+
"32k": 87.5,
|
| 965 |
+
"64k": 80.6,
|
| 966 |
+
"128k": 66.7
|
| 967 |
+
},
|
| 968 |
+
"ruler_long_score": {
|
| 969 |
+
"base": 91.85,
|
| 970 |
+
"per_length_lc": {
|
| 971 |
+
"16k": -0.0125,
|
| 972 |
+
"32k": -0.0474,
|
| 973 |
+
"64k": -0.1225,
|
| 974 |
+
"128k": -0.2738
|
| 975 |
+
},
|
| 976 |
+
"avg_lc": -0.114
|
| 977 |
+
},
|
| 978 |
+
"helmet": null,
|
| 979 |
+
"recipe_class": "longrope",
|
| 980 |
+
"params_b": 3.8,
|
| 981 |
+
"native_context_k": 4,
|
| 982 |
+
"source": "ruler"
|
| 983 |
+
},
|
| 984 |
+
"phi-3-medium-14-b": {
|
| 985 |
+
"display_name": "phi3-medium-14b",
|
| 986 |
+
"ruler_per_ctx": {
|
| 987 |
+
"4k": 93.3,
|
| 988 |
+
"8k": 93.2,
|
| 989 |
+
"16k": 91.1,
|
| 990 |
+
"32k": 86.8,
|
| 991 |
+
"64k": 78.6,
|
| 992 |
+
"128k": 46.1
|
| 993 |
+
},
|
| 994 |
+
"ruler_long_score": {
|
| 995 |
+
"base": 93.25,
|
| 996 |
+
"per_length_lc": {
|
| 997 |
+
"16k": -0.0231,
|
| 998 |
+
"32k": -0.0692,
|
| 999 |
+
"64k": -0.1571,
|
| 1000 |
+
"128k": -0.5056
|
| 1001 |
+
},
|
| 1002 |
+
"avg_lc": -0.1888
|
| 1003 |
+
},
|
| 1004 |
+
"helmet": null,
|
| 1005 |
+
"recipe_class": "longrope",
|
| 1006 |
+
"params_b": 14,
|
| 1007 |
+
"native_context_k": 4,
|
| 1008 |
+
"source": "ruler"
|
| 1009 |
+
},
|
| 1010 |
+
"gpt-4": {
|
| 1011 |
+
"display_name": "GPT-4",
|
| 1012 |
+
"ruler_per_ctx": null,
|
| 1013 |
+
"ruler_long_score": null,
|
| 1014 |
+
"helmet": {
|
| 1015 |
+
"overall": 39.2,
|
| 1016 |
+
"categories": {
|
| 1017 |
+
"Recall": 73.5,
|
| 1018 |
+
"RAG": 64.7,
|
| 1019 |
+
"Cite": 1.2,
|
| 1020 |
+
"Re-rank": 9.6,
|
| 1021 |
+
"ICL": 46.0,
|
| 1022 |
+
"LongQA": 46.7,
|
| 1023 |
+
"Summ": 33.0
|
| 1024 |
+
},
|
| 1025 |
+
"context_window_k": 128.0,
|
| 1026 |
+
"params_b": "?",
|
| 1027 |
+
"type": "♯"
|
| 1028 |
+
},
|
| 1029 |
+
"recipe_class": null,
|
| 1030 |
+
"params_b": "?",
|
| 1031 |
+
"native_context_k": null,
|
| 1032 |
+
"source": "helmet"
|
| 1033 |
+
},
|
| 1034 |
+
"gpt-4-o-mini": {
|
| 1035 |
+
"display_name": "GPT-4o-mini",
|
| 1036 |
+
"ruler_per_ctx": null,
|
| 1037 |
+
"ruler_long_score": null,
|
| 1038 |
+
"helmet": {
|
| 1039 |
+
"overall": 55.7,
|
| 1040 |
+
"categories": {
|
| 1041 |
+
"Recall": 89.6,
|
| 1042 |
+
"RAG": 68.0,
|
| 1043 |
+
"Cite": 24.3,
|
| 1044 |
+
"Re-rank": 31.2,
|
| 1045 |
+
"ICL": 81.2,
|
| 1046 |
+
"LongQA": 54.8,
|
| 1047 |
+
"Summ": 40.8
|
| 1048 |
+
},
|
| 1049 |
+
"context_window_k": 128.0,
|
| 1050 |
+
"params_b": "?",
|
| 1051 |
+
"type": "♯"
|
| 1052 |
+
},
|
| 1053 |
+
"recipe_class": null,
|
| 1054 |
+
"params_b": "?",
|
| 1055 |
+
"native_context_k": null,
|
| 1056 |
+
"source": "helmet"
|
| 1057 |
+
},
|
| 1058 |
+
"gpt-4-o-05": {
|
| 1059 |
+
"display_name": "GPT-4o-05",
|
| 1060 |
+
"ruler_per_ctx": null,
|
| 1061 |
+
"ruler_long_score": null,
|
| 1062 |
+
"helmet": {
|
| 1063 |
+
"overall": 56.4,
|
| 1064 |
+
"categories": {
|
| 1065 |
+
"Recall": 82.4,
|
| 1066 |
+
"RAG": 70.4,
|
| 1067 |
+
"Cite": 40.4,
|
| 1068 |
+
"Re-rank": 47.7,
|
| 1069 |
+
"ICL": 48.4,
|
| 1070 |
+
"LongQA": 62.0,
|
| 1071 |
+
"Summ": 43.6
|
| 1072 |
+
},
|
| 1073 |
+
"context_window_k": 128.0,
|
| 1074 |
+
"params_b": "?",
|
| 1075 |
+
"type": "♯"
|
| 1076 |
+
},
|
| 1077 |
+
"recipe_class": null,
|
| 1078 |
+
"params_b": "?",
|
| 1079 |
+
"native_context_k": null,
|
| 1080 |
+
"source": "helmet"
|
| 1081 |
+
},
|
| 1082 |
+
"gpt-4-o-08": {
|
| 1083 |
+
"display_name": "GPT-4o-08",
|
| 1084 |
+
"ruler_per_ctx": null,
|
| 1085 |
+
"ruler_long_score": null,
|
| 1086 |
+
"helmet": {
|
| 1087 |
+
"overall": 64.8,
|
| 1088 |
+
"categories": {
|
| 1089 |
+
"Recall": 99.9,
|
| 1090 |
+
"RAG": 70.2,
|
| 1091 |
+
"Cite": 44.3,
|
| 1092 |
+
"Re-rank": 50.0,
|
| 1093 |
+
"ICL": 86.3,
|
| 1094 |
+
"LongQA": 59.3,
|
| 1095 |
+
"Summ": 43.2
|
| 1096 |
+
},
|
| 1097 |
+
"context_window_k": 128.0,
|
| 1098 |
+
"params_b": "?",
|
| 1099 |
+
"type": "♯"
|
| 1100 |
+
},
|
| 1101 |
+
"recipe_class": null,
|
| 1102 |
+
"params_b": "?",
|
| 1103 |
+
"native_context_k": null,
|
| 1104 |
+
"source": "helmet"
|
| 1105 |
+
},
|
| 1106 |
+
"claude-3-5-sonnet": {
|
| 1107 |
+
"display_name": "Claude-3.5-Sonnet",
|
| 1108 |
+
"ruler_per_ctx": null,
|
| 1109 |
+
"ruler_long_score": null,
|
| 1110 |
+
"helmet": {
|
| 1111 |
+
"overall": 38.4,
|
| 1112 |
+
"categories": {
|
| 1113 |
+
"Recall": 94.7,
|
| 1114 |
+
"RAG": 38.1,
|
| 1115 |
+
"Cite": 18.7,
|
| 1116 |
+
"Re-rank": 7.2,
|
| 1117 |
+
"ICL": 61.0,
|
| 1118 |
+
"LongQA": 12.6,
|
| 1119 |
+
"Summ": 36.6
|
| 1120 |
+
},
|
| 1121 |
+
"context_window_k": 200.0,
|
| 1122 |
+
"params_b": "?",
|
| 1123 |
+
"type": "♯"
|
| 1124 |
+
},
|
| 1125 |
+
"recipe_class": null,
|
| 1126 |
+
"params_b": "?",
|
| 1127 |
+
"native_context_k": null,
|
| 1128 |
+
"source": "helmet"
|
| 1129 |
+
},
|
| 1130 |
+
"gemini-1-5-flash": {
|
| 1131 |
+
"display_name": "Gemini-1.5-Flash",
|
| 1132 |
+
"ruler_per_ctx": null,
|
| 1133 |
+
"ruler_long_score": null,
|
| 1134 |
+
"helmet": {
|
| 1135 |
+
"overall": 52.1,
|
| 1136 |
+
"categories": {
|
| 1137 |
+
"Recall": 91.2,
|
| 1138 |
+
"RAG": 66.5,
|
| 1139 |
+
"Cite": 31.7,
|
| 1140 |
+
"Re-rank": 50.8,
|
| 1141 |
+
"ICL": 28.1,
|
| 1142 |
+
"LongQA": 57.3,
|
| 1143 |
+
"Summ": 39.4
|
| 1144 |
+
},
|
| 1145 |
+
"context_window_k": 1024.0,
|
| 1146 |
+
"params_b": "?",
|
| 1147 |
+
"type": "♯"
|
| 1148 |
+
},
|
| 1149 |
+
"recipe_class": null,
|
| 1150 |
+
"params_b": "?",
|
| 1151 |
+
"native_context_k": null,
|
| 1152 |
+
"source": "helmet"
|
| 1153 |
+
},
|
| 1154 |
+
"gemini-1-5-pro": {
|
| 1155 |
+
"display_name": "Gemini-1.5-Pro",
|
| 1156 |
+
"ruler_per_ctx": null,
|
| 1157 |
+
"ruler_long_score": null,
|
| 1158 |
+
"helmet": {
|
| 1159 |
+
"overall": 64.4,
|
| 1160 |
+
"categories": {
|
| 1161 |
+
"Recall": 91.0,
|
| 1162 |
+
"RAG": 71.1,
|
| 1163 |
+
"Cite": 43.6,
|
| 1164 |
+
"Re-rank": 59.7,
|
| 1165 |
+
"ICL": 79.4,
|
| 1166 |
+
"LongQA": 59.6,
|
| 1167 |
+
"Summ": 46.4
|
| 1168 |
+
},
|
| 1169 |
+
"context_window_k": 2048.0,
|
| 1170 |
+
"params_b": "?",
|
| 1171 |
+
"type": "♯"
|
| 1172 |
+
},
|
| 1173 |
+
"recipe_class": null,
|
| 1174 |
+
"params_b": "?",
|
| 1175 |
+
"native_context_k": null,
|
| 1176 |
+
"source": "helmet"
|
| 1177 |
+
},
|
| 1178 |
+
"llama-2-7-b-32-k": {
|
| 1179 |
+
"display_name": "Llama-2-7B-32k",
|
| 1180 |
+
"ruler_per_ctx": null,
|
| 1181 |
+
"ruler_long_score": null,
|
| 1182 |
+
"helmet": {
|
| 1183 |
+
"overall": 1.3,
|
| 1184 |
+
"categories": {
|
| 1185 |
+
"Recall": 0.0,
|
| 1186 |
+
"RAG": 0.0,
|
| 1187 |
+
"Cite": 0.0,
|
| 1188 |
+
"Re-rank": 0.0,
|
| 1189 |
+
"ICL": 0.0,
|
| 1190 |
+
"LongQA": 9.4,
|
| 1191 |
+
"Summ": 0.0
|
| 1192 |
+
},
|
| 1193 |
+
"context_window_k": 32.0,
|
| 1194 |
+
"params_b": "7.0",
|
| 1195 |
+
"type": "♭"
|
| 1196 |
+
},
|
| 1197 |
+
"recipe_class": null,
|
| 1198 |
+
"params_b": "7.0",
|
| 1199 |
+
"native_context_k": null,
|
| 1200 |
+
"source": "helmet"
|
| 1201 |
+
},
|
| 1202 |
+
"llama-2-7-b-32-k-instruct": {
|
| 1203 |
+
"display_name": "Llama-2-7B-32k-Inst",
|
| 1204 |
+
"ruler_per_ctx": null,
|
| 1205 |
+
"ruler_long_score": null,
|
| 1206 |
+
"helmet": {
|
| 1207 |
+
"overall": 1.3,
|
| 1208 |
+
"categories": {
|
| 1209 |
+
"Recall": 0.0,
|
| 1210 |
+
"RAG": 0.0,
|
| 1211 |
+
"Cite": 0.0,
|
| 1212 |
+
"Re-rank": 0.0,
|
| 1213 |
+
"ICL": 0.0,
|
| 1214 |
+
"LongQA": 9.1,
|
| 1215 |
+
"Summ": 0.0
|
| 1216 |
+
},
|
| 1217 |
+
"context_window_k": 32.0,
|
| 1218 |
+
"params_b": "7.0",
|
| 1219 |
+
"type": "♯"
|
| 1220 |
+
},
|
| 1221 |
+
"recipe_class": null,
|
| 1222 |
+
"params_b": "7.0",
|
| 1223 |
+
"native_context_k": null,
|
| 1224 |
+
"source": "helmet"
|
| 1225 |
+
},
|
| 1226 |
+
"llama-7-b-80-k": {
|
| 1227 |
+
"display_name": "Llama-7B-80k",
|
| 1228 |
+
"ruler_per_ctx": null,
|
| 1229 |
+
"ruler_long_score": null,
|
| 1230 |
+
"helmet": {
|
| 1231 |
+
"overall": 22.4,
|
| 1232 |
+
"categories": {
|
| 1233 |
+
"Recall": 17.0,
|
| 1234 |
+
"RAG": 40.8,
|
| 1235 |
+
"Cite": 1.0,
|
| 1236 |
+
"Re-rank": 0.2,
|
| 1237 |
+
"ICL": 76.8,
|
| 1238 |
+
"LongQA": 19.8,
|
| 1239 |
+
"Summ": 0.9
|
| 1240 |
+
},
|
| 1241 |
+
"context_window_k": 80.0,
|
| 1242 |
+
"params_b": "7.0",
|
| 1243 |
+
"type": "♭"
|
| 1244 |
+
},
|
| 1245 |
+
"recipe_class": null,
|
| 1246 |
+
"params_b": "7.0",
|
| 1247 |
+
"native_context_k": null,
|
| 1248 |
+
"source": "helmet"
|
| 1249 |
+
},
|
| 1250 |
+
"yarn-llama-2-7-b-64-k": {
|
| 1251 |
+
"display_name": "Yarn-Llama-2-7B-64k",
|
| 1252 |
+
"ruler_per_ctx": null,
|
| 1253 |
+
"ruler_long_score": null,
|
| 1254 |
+
"helmet": {
|
| 1255 |
+
"overall": 0.6,
|
| 1256 |
+
"categories": {
|
| 1257 |
+
"Recall": 0.0,
|
| 1258 |
+
"RAG": 0.0,
|
| 1259 |
+
"Cite": 0.2,
|
| 1260 |
+
"Re-rank": 0.0,
|
| 1261 |
+
"ICL": 0.0,
|
| 1262 |
+
"LongQA": 3.8,
|
| 1263 |
+
"Summ": 0.2
|
| 1264 |
+
},
|
| 1265 |
+
"context_window_k": 64.0,
|
| 1266 |
+
"params_b": "7.0",
|
| 1267 |
+
"type": "♭"
|
| 1268 |
+
},
|
| 1269 |
+
"recipe_class": null,
|
| 1270 |
+
"params_b": "7.0",
|
| 1271 |
+
"native_context_k": null,
|
| 1272 |
+
"source": "helmet"
|
| 1273 |
+
},
|
| 1274 |
+
"yarn-llama-2-7-b-128-k": {
|
| 1275 |
+
"display_name": "Yarn-Llama-2-7B-128k",
|
| 1276 |
+
"ruler_per_ctx": null,
|
| 1277 |
+
"ruler_long_score": null,
|
| 1278 |
+
"helmet": {
|
| 1279 |
+
"overall": 17.9,
|
| 1280 |
+
"categories": {
|
| 1281 |
+
"Recall": 4.3,
|
| 1282 |
+
"RAG": 26.7,
|
| 1283 |
+
"Cite": 0.7,
|
| 1284 |
+
"Re-rank": 0.4,
|
| 1285 |
+
"ICL": 75.4,
|
| 1286 |
+
"LongQA": 16.1,
|
| 1287 |
+
"Summ": 1.9
|
| 1288 |
+
},
|
| 1289 |
+
"context_window_k": 128.0,
|
| 1290 |
+
"params_b": "7.0",
|
| 1291 |
+
"type": "♭"
|
| 1292 |
+
},
|
| 1293 |
+
"recipe_class": null,
|
| 1294 |
+
"params_b": "7.0",
|
| 1295 |
+
"native_context_k": null,
|
| 1296 |
+
"source": "helmet"
|
| 1297 |
+
},
|
| 1298 |
+
"llama-3-8-b": {
|
| 1299 |
+
"display_name": "Llama-3-8B",
|
| 1300 |
+
"ruler_per_ctx": null,
|
| 1301 |
+
"ruler_long_score": null,
|
| 1302 |
+
"helmet": {
|
| 1303 |
+
"overall": 1.5,
|
| 1304 |
+
"categories": {
|
| 1305 |
+
"Recall": 0.0,
|
| 1306 |
+
"RAG": 0.0,
|
| 1307 |
+
"Cite": 0.0,
|
| 1308 |
+
"Re-rank": 0.0,
|
| 1309 |
+
"ICL": 0.0,
|
| 1310 |
+
"LongQA": 10.5,
|
| 1311 |
+
"Summ": 0.0
|
| 1312 |
+
},
|
| 1313 |
+
"context_window_k": 8.0,
|
| 1314 |
+
"params_b": "8.0",
|
| 1315 |
+
"type": "♭"
|
| 1316 |
+
},
|
| 1317 |
+
"recipe_class": null,
|
| 1318 |
+
"params_b": "8.0",
|
| 1319 |
+
"native_context_k": null,
|
| 1320 |
+
"source": "helmet"
|
| 1321 |
+
},
|
| 1322 |
+
"llama-3-8-b-instruct": {
|
| 1323 |
+
"display_name": "Llama-3-8B-Inst",
|
| 1324 |
+
"ruler_per_ctx": null,
|
| 1325 |
+
"ruler_long_score": null,
|
| 1326 |
+
"helmet": {
|
| 1327 |
+
"overall": 1.2,
|
| 1328 |
+
"categories": {
|
| 1329 |
+
"Recall": 0.0,
|
| 1330 |
+
"RAG": 0.0,
|
| 1331 |
+
"Cite": 0.0,
|
| 1332 |
+
"Re-rank": 0.0,
|
| 1333 |
+
"ICL": 0.0,
|
| 1334 |
+
"LongQA": 8.7,
|
| 1335 |
+
"Summ": 0.0
|
| 1336 |
+
},
|
| 1337 |
+
"context_window_k": 8.0,
|
| 1338 |
+
"params_b": "8.0",
|
| 1339 |
+
"type": "♯"
|
| 1340 |
+
},
|
| 1341 |
+
"recipe_class": null,
|
| 1342 |
+
"params_b": "8.0",
|
| 1343 |
+
"native_context_k": null,
|
| 1344 |
+
"source": "helmet"
|
| 1345 |
+
},
|
| 1346 |
+
"llama-3-8-b-θ": {
|
| 1347 |
+
"display_name": "Llama-3-8B-θ",
|
| 1348 |
+
"ruler_per_ctx": null,
|
| 1349 |
+
"ruler_long_score": null,
|
| 1350 |
+
"helmet": {
|
| 1351 |
+
"overall": 0.9,
|
| 1352 |
+
"categories": {
|
| 1353 |
+
"Recall": 0.0,
|
| 1354 |
+
"RAG": 2.4,
|
| 1355 |
+
"Cite": 0.1,
|
| 1356 |
+
"Re-rank": 0.0,
|
| 1357 |
+
"ICL": 2.8,
|
| 1358 |
+
"LongQA": 0.8,
|
| 1359 |
+
"Summ": 0.0
|
| 1360 |
+
},
|
| 1361 |
+
"context_window_k": 8.0,
|
| 1362 |
+
"params_b": "8.0",
|
| 1363 |
+
"type": "♭"
|
| 1364 |
+
},
|
| 1365 |
+
"recipe_class": null,
|
| 1366 |
+
"params_b": "8.0",
|
| 1367 |
+
"native_context_k": null,
|
| 1368 |
+
"source": "helmet"
|
| 1369 |
+
},
|
| 1370 |
+
"llama-3-8-b-instruct-θ": {
|
| 1371 |
+
"display_name": "Llama-3-8B-Inst-θ",
|
| 1372 |
+
"ruler_per_ctx": null,
|
| 1373 |
+
"ruler_long_score": null,
|
| 1374 |
+
"helmet": {
|
| 1375 |
+
"overall": 1.6,
|
| 1376 |
+
"categories": {
|
| 1377 |
+
"Recall": 0.0,
|
| 1378 |
+
"RAG": 1.3,
|
| 1379 |
+
"Cite": 0.8,
|
| 1380 |
+
"Re-rank": 0.1,
|
| 1381 |
+
"ICL": 4.3,
|
| 1382 |
+
"LongQA": 2.5,
|
| 1383 |
+
"Summ": 1.9
|
| 1384 |
+
},
|
| 1385 |
+
"context_window_k": 8.0,
|
| 1386 |
+
"params_b": "8.0",
|
| 1387 |
+
"type": "♯"
|
| 1388 |
+
},
|
| 1389 |
+
"recipe_class": null,
|
| 1390 |
+
"params_b": "8.0",
|
| 1391 |
+
"native_context_k": null,
|
| 1392 |
+
"source": "helmet"
|
| 1393 |
+
},
|
| 1394 |
+
"llama-3-70-b-θ": {
|
| 1395 |
+
"display_name": "Llama-3-70B-θ",
|
| 1396 |
+
"ruler_per_ctx": null,
|
| 1397 |
+
"ruler_long_score": null,
|
| 1398 |
+
"helmet": {
|
| 1399 |
+
"overall": 1.2,
|
| 1400 |
+
"categories": {
|
| 1401 |
+
"Recall": 0.0,
|
| 1402 |
+
"RAG": 0.0,
|
| 1403 |
+
"Cite": 2.9,
|
| 1404 |
+
"Re-rank": 0.0,
|
| 1405 |
+
"ICL": 0.2,
|
| 1406 |
+
"LongQA": 5.4,
|
| 1407 |
+
"Summ": 0.1
|
| 1408 |
+
},
|
| 1409 |
+
"context_window_k": 8.0,
|
| 1410 |
+
"params_b": "70.0",
|
| 1411 |
+
"type": "♭"
|
| 1412 |
+
},
|
| 1413 |
+
"recipe_class": null,
|
| 1414 |
+
"params_b": "70.0",
|
| 1415 |
+
"native_context_k": null,
|
| 1416 |
+
"source": "helmet"
|
| 1417 |
+
},
|
| 1418 |
+
"llama-3-70-b-instruct-θ": {
|
| 1419 |
+
"display_name": "Llama-3-70B-Inst-θ",
|
| 1420 |
+
"ruler_per_ctx": null,
|
| 1421 |
+
"ruler_long_score": null,
|
| 1422 |
+
"helmet": {
|
| 1423 |
+
"overall": 2.1,
|
| 1424 |
+
"categories": {
|
| 1425 |
+
"Recall": 0.0,
|
| 1426 |
+
"RAG": 0.0,
|
| 1427 |
+
"Cite": 0.0,
|
| 1428 |
+
"Re-rank": 0.0,
|
| 1429 |
+
"ICL": 0.0,
|
| 1430 |
+
"LongQA": 12.0,
|
| 1431 |
+
"Summ": 2.7
|
| 1432 |
+
},
|
| 1433 |
+
"context_window_k": 8.0,
|
| 1434 |
+
"params_b": "70.0",
|
| 1435 |
+
"type": "♯"
|
| 1436 |
+
},
|
| 1437 |
+
"recipe_class": null,
|
| 1438 |
+
"params_b": "70.0",
|
| 1439 |
+
"native_context_k": null,
|
| 1440 |
+
"source": "helmet"
|
| 1441 |
+
},
|
| 1442 |
+
"llama-3-1-8-b": {
|
| 1443 |
+
"display_name": "Llama-3.1-8B",
|
| 1444 |
+
"ruler_per_ctx": null,
|
| 1445 |
+
"ruler_long_score": null,
|
| 1446 |
+
"helmet": {
|
| 1447 |
+
"overall": 37.1,
|
| 1448 |
+
"categories": {
|
| 1449 |
+
"Recall": 76.6,
|
| 1450 |
+
"RAG": 54.5,
|
| 1451 |
+
"Cite": 1.6,
|
| 1452 |
+
"Re-rank": 9.3,
|
| 1453 |
+
"ICL": 79.5,
|
| 1454 |
+
"LongQA": 36.2,
|
| 1455 |
+
"Summ": 1.7
|
| 1456 |
+
},
|
| 1457 |
+
"context_window_k": 128.0,
|
| 1458 |
+
"params_b": "8.0",
|
| 1459 |
+
"type": "♭"
|
| 1460 |
+
},
|
| 1461 |
+
"recipe_class": null,
|
| 1462 |
+
"params_b": "8.0",
|
| 1463 |
+
"native_context_k": null,
|
| 1464 |
+
"source": "helmet"
|
| 1465 |
+
},
|
| 1466 |
+
"llama-3-1-70-b": {
|
| 1467 |
+
"display_name": "Llama-3.1-70B",
|
| 1468 |
+
"ruler_per_ctx": null,
|
| 1469 |
+
"ruler_long_score": null,
|
| 1470 |
+
"helmet": {
|
| 1471 |
+
"overall": 41.7,
|
| 1472 |
+
"categories": {
|
| 1473 |
+
"Recall": 76.3,
|
| 1474 |
+
"RAG": 56.6,
|
| 1475 |
+
"Cite": 3.9,
|
| 1476 |
+
"Re-rank": 22.6,
|
| 1477 |
+
"ICL": 80.4,
|
| 1478 |
+
"LongQA": 46.0,
|
| 1479 |
+
"Summ": 6.3
|
| 1480 |
+
},
|
| 1481 |
+
"context_window_k": 128.0,
|
| 1482 |
+
"params_b": "70.0",
|
| 1483 |
+
"type": "♭"
|
| 1484 |
+
},
|
| 1485 |
+
"recipe_class": null,
|
| 1486 |
+
"params_b": "70.0",
|
| 1487 |
+
"native_context_k": null,
|
| 1488 |
+
"source": "helmet"
|
| 1489 |
+
},
|
| 1490 |
+
"llama-3-3-70-b-instruct": {
|
| 1491 |
+
"display_name": "Llama-3.3-70B-Inst",
|
| 1492 |
+
"ruler_per_ctx": null,
|
| 1493 |
+
"ruler_long_score": null,
|
| 1494 |
+
"helmet": {
|
| 1495 |
+
"overall": 48.2,
|
| 1496 |
+
"categories": {
|
| 1497 |
+
"Recall": 81.8,
|
| 1498 |
+
"RAG": 55.1,
|
| 1499 |
+
"Cite": 10.9,
|
| 1500 |
+
"Re-rank": 26.2,
|
| 1501 |
+
"ICL": 77.8,
|
| 1502 |
+
"LongQA": 52.0,
|
| 1503 |
+
"Summ": 33.3
|
| 1504 |
+
},
|
| 1505 |
+
"context_window_k": 128.0,
|
| 1506 |
+
"params_b": "70.0",
|
| 1507 |
+
"type": "♯"
|
| 1508 |
+
},
|
| 1509 |
+
"recipe_class": null,
|
| 1510 |
+
"params_b": "70.0",
|
| 1511 |
+
"native_context_k": null,
|
| 1512 |
+
"source": "helmet"
|
| 1513 |
+
},
|
| 1514 |
+
"llama-3-2-1-b": {
|
| 1515 |
+
"display_name": "Llama-3.2-1B",
|
| 1516 |
+
"ruler_per_ctx": null,
|
| 1517 |
+
"ruler_long_score": null,
|
| 1518 |
+
"helmet": {
|
| 1519 |
+
"overall": 22.6,
|
| 1520 |
+
"categories": {
|
| 1521 |
+
"Recall": 25.2,
|
| 1522 |
+
"RAG": 32.1,
|
| 1523 |
+
"Cite": 1.5,
|
| 1524 |
+
"Re-rank": 5.6,
|
| 1525 |
+
"ICL": 76.9,
|
| 1526 |
+
"LongQA": 15.9,
|
| 1527 |
+
"Summ": 0.9
|
| 1528 |
+
},
|
| 1529 |
+
"context_window_k": 128.0,
|
| 1530 |
+
"params_b": "1.0",
|
| 1531 |
+
"type": "♭"
|
| 1532 |
+
},
|
| 1533 |
+
"recipe_class": null,
|
| 1534 |
+
"params_b": "1.0",
|
| 1535 |
+
"native_context_k": null,
|
| 1536 |
+
"source": "helmet"
|
| 1537 |
+
},
|
| 1538 |
+
"llama-3-2-1-b-instruct": {
|
| 1539 |
+
"display_name": "Llama-3.2-1B-Inst",
|
| 1540 |
+
"ruler_per_ctx": null,
|
| 1541 |
+
"ruler_long_score": null,
|
| 1542 |
+
"helmet": {
|
| 1543 |
+
"overall": 25.0,
|
| 1544 |
+
"categories": {
|
| 1545 |
+
"Recall": 32.4,
|
| 1546 |
+
"RAG": 33.8,
|
| 1547 |
+
"Cite": 2.4,
|
| 1548 |
+
"Re-rank": 1.7,
|
| 1549 |
+
"ICL": 77.3,
|
| 1550 |
+
"LongQA": 18.1,
|
| 1551 |
+
"Summ": 9.1
|
| 1552 |
+
},
|
| 1553 |
+
"context_window_k": 128.0,
|
| 1554 |
+
"params_b": "1.0",
|
| 1555 |
+
"type": "♯"
|
| 1556 |
+
},
|
| 1557 |
+
"recipe_class": null,
|
| 1558 |
+
"params_b": "1.0",
|
| 1559 |
+
"native_context_k": null,
|
| 1560 |
+
"source": "helmet"
|
| 1561 |
+
},
|
| 1562 |
+
"llama-3-2-3-b": {
|
| 1563 |
+
"display_name": "Llama-3.2-3B",
|
| 1564 |
+
"ruler_per_ctx": null,
|
| 1565 |
+
"ruler_long_score": null,
|
| 1566 |
+
"helmet": {
|
| 1567 |
+
"overall": 33.5,
|
| 1568 |
+
"categories": {
|
| 1569 |
+
"Recall": 58.8,
|
| 1570 |
+
"RAG": 51.4,
|
| 1571 |
+
"Cite": 2.1,
|
| 1572 |
+
"Re-rank": 5.4,
|
| 1573 |
+
"ICL": 89.5,
|
| 1574 |
+
"LongQA": 27.6,
|
| 1575 |
+
"Summ": 0.0
|
| 1576 |
+
},
|
| 1577 |
+
"context_window_k": 128.0,
|
| 1578 |
+
"params_b": "3.0",
|
| 1579 |
+
"type": "♭"
|
| 1580 |
+
},
|
| 1581 |
+
"recipe_class": null,
|
| 1582 |
+
"params_b": "3.0",
|
| 1583 |
+
"native_context_k": null,
|
| 1584 |
+
"source": "helmet"
|
| 1585 |
+
},
|
| 1586 |
+
"llama-3-2-3-b-instruct": {
|
| 1587 |
+
"display_name": "Llama-3.2-3B-Inst",
|
| 1588 |
+
"ruler_per_ctx": null,
|
| 1589 |
+
"ruler_long_score": null,
|
| 1590 |
+
"helmet": {
|
| 1591 |
+
"overall": 37.0,
|
| 1592 |
+
"categories": {
|
| 1593 |
+
"Recall": 53.1,
|
| 1594 |
+
"RAG": 56.4,
|
| 1595 |
+
"Cite": 7.4,
|
| 1596 |
+
"Re-rank": 0.7,
|
| 1597 |
+
"ICL": 86.4,
|
| 1598 |
+
"LongQA": 28.9,
|
| 1599 |
+
"Summ": 25.9
|
| 1600 |
+
},
|
| 1601 |
+
"context_window_k": 128.0,
|
| 1602 |
+
"params_b": "3.0",
|
| 1603 |
+
"type": "♯"
|
| 1604 |
+
},
|
| 1605 |
+
"recipe_class": null,
|
| 1606 |
+
"params_b": "3.0",
|
| 1607 |
+
"native_context_k": null,
|
| 1608 |
+
"source": "helmet"
|
| 1609 |
+
},
|
| 1610 |
+
"llama-4-17-b": {
|
| 1611 |
+
"display_name": "Llama-4-17B",
|
| 1612 |
+
"ruler_per_ctx": null,
|
| 1613 |
+
"ruler_long_score": null,
|
| 1614 |
+
"helmet": {
|
| 1615 |
+
"overall": 31.4,
|
| 1616 |
+
"categories": {
|
| 1617 |
+
"Recall": 42.8,
|
| 1618 |
+
"RAG": 50.4,
|
| 1619 |
+
"Cite": 3.7,
|
| 1620 |
+
"Re-rank": 6.1,
|
| 1621 |
+
"ICL": 88.0,
|
| 1622 |
+
"LongQA": 28.1,
|
| 1623 |
+
"Summ": 0.6
|
| 1624 |
+
},
|
| 1625 |
+
"context_window_k": 128.0,
|
| 1626 |
+
"params_b": "17.0",
|
| 1627 |
+
"type": "♭"
|
| 1628 |
+
},
|
| 1629 |
+
"recipe_class": null,
|
| 1630 |
+
"params_b": "17.0",
|
| 1631 |
+
"native_context_k": null,
|
| 1632 |
+
"source": "helmet"
|
| 1633 |
+
},
|
| 1634 |
+
"llama-4-17-b-instruct": {
|
| 1635 |
+
"display_name": "Llama-4-17B-Inst",
|
| 1636 |
+
"ruler_per_ctx": null,
|
| 1637 |
+
"ruler_long_score": null,
|
| 1638 |
+
"helmet": {
|
| 1639 |
+
"overall": 45.0,
|
| 1640 |
+
"categories": {
|
| 1641 |
+
"Recall": 76.6,
|
| 1642 |
+
"RAG": 51.2,
|
| 1643 |
+
"Cite": 4.7,
|
| 1644 |
+
"Re-rank": 17.9,
|
| 1645 |
+
"ICL": 81.0,
|
| 1646 |
+
"LongQA": 50.5,
|
| 1647 |
+
"Summ": 33.2
|
| 1648 |
+
},
|
| 1649 |
+
"context_window_k": 128.0,
|
| 1650 |
+
"params_b": "17.0",
|
| 1651 |
+
"type": "♯"
|
| 1652 |
+
},
|
| 1653 |
+
"recipe_class": null,
|
| 1654 |
+
"params_b": "17.0",
|
| 1655 |
+
"native_context_k": null,
|
| 1656 |
+
"source": "helmet"
|
| 1657 |
+
},
|
| 1658 |
+
"mistral-7-b-instruct-v-0-1": {
|
| 1659 |
+
"display_name": "Mistral-7B-Inst-v0.1",
|
| 1660 |
+
"ruler_per_ctx": null,
|
| 1661 |
+
"ruler_long_score": null,
|
| 1662 |
+
"helmet": {
|
| 1663 |
+
"overall": 15.8,
|
| 1664 |
+
"categories": {
|
| 1665 |
+
"Recall": 3.8,
|
| 1666 |
+
"RAG": 36.6,
|
| 1667 |
+
"Cite": 1.3,
|
| 1668 |
+
"Re-rank": 0.0,
|
| 1669 |
+
"ICL": 44.8,
|
| 1670 |
+
"LongQA": 15.2,
|
| 1671 |
+
"Summ": 9.2
|
| 1672 |
+
},
|
| 1673 |
+
"context_window_k": 8.0,
|
| 1674 |
+
"params_b": "7.0",
|
| 1675 |
+
"type": "♯"
|
| 1676 |
+
},
|
| 1677 |
+
"recipe_class": null,
|
| 1678 |
+
"params_b": "7.0",
|
| 1679 |
+
"native_context_k": null,
|
| 1680 |
+
"source": "helmet"
|
| 1681 |
+
},
|
| 1682 |
+
"mistral-7-b-instruct-v-0-2": {
|
| 1683 |
+
"display_name": "Mistral-7B-Inst-v0.2",
|
| 1684 |
+
"ruler_per_ctx": null,
|
| 1685 |
+
"ruler_long_score": null,
|
| 1686 |
+
"helmet": {
|
| 1687 |
+
"overall": 8.7,
|
| 1688 |
+
"categories": {
|
| 1689 |
+
"Recall": 1.5,
|
| 1690 |
+
"RAG": 11.4,
|
| 1691 |
+
"Cite": 0.6,
|
| 1692 |
+
"Re-rank": 1.2,
|
| 1693 |
+
"ICL": 38.0,
|
| 1694 |
+
"LongQA": 4.1,
|
| 1695 |
+
"Summ": 4.3
|
| 1696 |
+
},
|
| 1697 |
+
"context_window_k": 32.0,
|
| 1698 |
+
"params_b": "7.0",
|
| 1699 |
+
"type": "♯"
|
| 1700 |
+
},
|
| 1701 |
+
"recipe_class": null,
|
| 1702 |
+
"params_b": "7.0",
|
| 1703 |
+
"native_context_k": null,
|
| 1704 |
+
"source": "helmet"
|
| 1705 |
+
},
|
| 1706 |
+
"mistral-7-b-v-0-3": {
|
| 1707 |
+
"display_name": "Mistral-7B-v0.3",
|
| 1708 |
+
"ruler_per_ctx": null,
|
| 1709 |
+
"ruler_long_score": null,
|
| 1710 |
+
"helmet": {
|
| 1711 |
+
"overall": 10.8,
|
| 1712 |
+
"categories": {
|
| 1713 |
+
"Recall": 3.6,
|
| 1714 |
+
"RAG": 4.3,
|
| 1715 |
+
"Cite": 0.4,
|
| 1716 |
+
"Re-rank": 1.2,
|
| 1717 |
+
"ICL": 54.8,
|
| 1718 |
+
"LongQA": 10.7,
|
| 1719 |
+
"Summ": 0.4
|
| 1720 |
+
},
|
| 1721 |
+
"context_window_k": 32.0,
|
| 1722 |
+
"params_b": "7.0",
|
| 1723 |
+
"type": "♭"
|
| 1724 |
+
},
|
| 1725 |
+
"recipe_class": null,
|
| 1726 |
+
"params_b": "7.0",
|
| 1727 |
+
"native_context_k": null,
|
| 1728 |
+
"source": "helmet"
|
| 1729 |
+
},
|
| 1730 |
+
"mistral-7-b-instruct-v-0-3": {
|
| 1731 |
+
"display_name": "Mistral-7B-Inst-v0.3",
|
| 1732 |
+
"ruler_per_ctx": null,
|
| 1733 |
+
"ruler_long_score": null,
|
| 1734 |
+
"helmet": {
|
| 1735 |
+
"overall": 17.3,
|
| 1736 |
+
"categories": {
|
| 1737 |
+
"Recall": 8.0,
|
| 1738 |
+
"RAG": 21.3,
|
| 1739 |
+
"Cite": 0.9,
|
| 1740 |
+
"Re-rank": 0.5,
|
| 1741 |
+
"ICL": 73.6,
|
| 1742 |
+
"LongQA": 12.0,
|
| 1743 |
+
"Summ": 5.1
|
| 1744 |
+
},
|
| 1745 |
+
"context_window_k": 32.0,
|
| 1746 |
+
"params_b": "7.0",
|
| 1747 |
+
"type": "♯"
|
| 1748 |
+
},
|
| 1749 |
+
"recipe_class": null,
|
| 1750 |
+
"params_b": "7.0",
|
| 1751 |
+
"native_context_k": null,
|
| 1752 |
+
"source": "helmet"
|
| 1753 |
+
},
|
| 1754 |
+
"ministral-8-b-instruct": {
|
| 1755 |
+
"display_name": "Ministral-8B-Inst",
|
| 1756 |
+
"ruler_per_ctx": null,
|
| 1757 |
+
"ruler_long_score": null,
|
| 1758 |
+
"helmet": {
|
| 1759 |
+
"overall": 24.2,
|
| 1760 |
+
"categories": {
|
| 1761 |
+
"Recall": 12.8,
|
| 1762 |
+
"RAG": 35.3,
|
| 1763 |
+
"Cite": 0.4,
|
| 1764 |
+
"Re-rank": 0.0,
|
| 1765 |
+
"ICL": 84.4,
|
| 1766 |
+
"LongQA": 25.0,
|
| 1767 |
+
"Summ": 11.5
|
| 1768 |
+
},
|
| 1769 |
+
"context_window_k": 128.0,
|
| 1770 |
+
"params_b": "8.0",
|
| 1771 |
+
"type": "♯"
|
| 1772 |
+
},
|
| 1773 |
+
"recipe_class": null,
|
| 1774 |
+
"params_b": "8.0",
|
| 1775 |
+
"native_context_k": null,
|
| 1776 |
+
"source": "helmet"
|
| 1777 |
+
},
|
| 1778 |
+
"mistral-nemo": {
|
| 1779 |
+
"display_name": "Mistral-Nemo",
|
| 1780 |
+
"ruler_per_ctx": null,
|
| 1781 |
+
"ruler_long_score": null,
|
| 1782 |
+
"helmet": {
|
| 1783 |
+
"overall": 25.0,
|
| 1784 |
+
"categories": {
|
| 1785 |
+
"Recall": 17.7,
|
| 1786 |
+
"RAG": 45.9,
|
| 1787 |
+
"Cite": 1.0,
|
| 1788 |
+
"Re-rank": 0.0,
|
| 1789 |
+
"ICL": 82.4,
|
| 1790 |
+
"LongQA": 26.2,
|
| 1791 |
+
"Summ": 1.7
|
| 1792 |
+
},
|
| 1793 |
+
"context_window_k": 128.0,
|
| 1794 |
+
"params_b": "12.0",
|
| 1795 |
+
"type": "♭"
|
| 1796 |
+
},
|
| 1797 |
+
"recipe_class": null,
|
| 1798 |
+
"params_b": "12.0",
|
| 1799 |
+
"native_context_k": null,
|
| 1800 |
+
"source": "helmet"
|
| 1801 |
+
},
|
| 1802 |
+
"mistral-nemo-instruct": {
|
| 1803 |
+
"display_name": "Mistral-Nemo-Inst",
|
| 1804 |
+
"ruler_per_ctx": null,
|
| 1805 |
+
"ruler_long_score": null,
|
| 1806 |
+
"helmet": {
|
| 1807 |
+
"overall": 25.7,
|
| 1808 |
+
"categories": {
|
| 1809 |
+
"Recall": 14.6,
|
| 1810 |
+
"RAG": 40.0,
|
| 1811 |
+
"Cite": 0.5,
|
| 1812 |
+
"Re-rank": 0.0,
|
| 1813 |
+
"ICL": 84.0,
|
| 1814 |
+
"LongQA": 22.5,
|
| 1815 |
+
"Summ": 18.5
|
| 1816 |
+
},
|
| 1817 |
+
"context_window_k": 128.0,
|
| 1818 |
+
"params_b": "12.0",
|
| 1819 |
+
"type": "♯"
|
| 1820 |
+
},
|
| 1821 |
+
"recipe_class": null,
|
| 1822 |
+
"params_b": "12.0",
|
| 1823 |
+
"native_context_k": null,
|
| 1824 |
+
"source": "helmet"
|
| 1825 |
+
},
|
| 1826 |
+
"megabeam-mistral": {
|
| 1827 |
+
"display_name": "MegaBeam-Mistral",
|
| 1828 |
+
"ruler_per_ctx": null,
|
| 1829 |
+
"ruler_long_score": null,
|
| 1830 |
+
"helmet": {
|
| 1831 |
+
"overall": 45.4,
|
| 1832 |
+
"categories": {
|
| 1833 |
+
"Recall": 89.6,
|
| 1834 |
+
"RAG": 57.0,
|
| 1835 |
+
"Cite": 4.0,
|
| 1836 |
+
"Re-rank": 14.7,
|
| 1837 |
+
"ICL": 86.2,
|
| 1838 |
+
"LongQA": 37.3,
|
| 1839 |
+
"Summ": 28.9
|
| 1840 |
+
},
|
| 1841 |
+
"context_window_k": 512.0,
|
| 1842 |
+
"params_b": "7.0",
|
| 1843 |
+
"type": "♯"
|
| 1844 |
+
},
|
| 1845 |
+
"recipe_class": null,
|
| 1846 |
+
"params_b": "7.0",
|
| 1847 |
+
"native_context_k": null,
|
| 1848 |
+
"source": "helmet"
|
| 1849 |
+
},
|
| 1850 |
+
"yi-6-b-200-k": {
|
| 1851 |
+
"display_name": "Yi-6B-200k",
|
| 1852 |
+
"ruler_per_ctx": null,
|
| 1853 |
+
"ruler_long_score": null,
|
| 1854 |
+
"helmet": {
|
| 1855 |
+
"overall": 27.8,
|
| 1856 |
+
"categories": {
|
| 1857 |
+
"Recall": 37.4,
|
| 1858 |
+
"RAG": 36.9,
|
| 1859 |
+
"Cite": 1.1,
|
| 1860 |
+
"Re-rank": 3.4,
|
| 1861 |
+
"ICL": 82.3,
|
| 1862 |
+
"LongQA": 30.0,
|
| 1863 |
+
"Summ": 3.5
|
| 1864 |
+
},
|
| 1865 |
+
"context_window_k": 200.0,
|
| 1866 |
+
"params_b": "6.0",
|
| 1867 |
+
"type": "♭"
|
| 1868 |
+
},
|
| 1869 |
+
"recipe_class": null,
|
| 1870 |
+
"params_b": "6.0",
|
| 1871 |
+
"native_context_k": null,
|
| 1872 |
+
"source": "helmet"
|
| 1873 |
+
},
|
| 1874 |
+
"yi-9-b-200-k": {
|
| 1875 |
+
"display_name": "Yi-9B-200k",
|
| 1876 |
+
"ruler_per_ctx": null,
|
| 1877 |
+
"ruler_long_score": null,
|
| 1878 |
+
"helmet": {
|
| 1879 |
+
"overall": 33.9,
|
| 1880 |
+
"categories": {
|
| 1881 |
+
"Recall": 56.2,
|
| 1882 |
+
"RAG": 50.5,
|
| 1883 |
+
"Cite": 3.3,
|
| 1884 |
+
"Re-rank": 8.3,
|
| 1885 |
+
"ICL": 73.4,
|
| 1886 |
+
"LongQA": 39.3,
|
| 1887 |
+
"Summ": 6.2
|
| 1888 |
+
},
|
| 1889 |
+
"context_window_k": 200.0,
|
| 1890 |
+
"params_b": "9.0",
|
| 1891 |
+
"type": "♭"
|
| 1892 |
+
},
|
| 1893 |
+
"recipe_class": null,
|
| 1894 |
+
"params_b": "9.0",
|
| 1895 |
+
"native_context_k": null,
|
| 1896 |
+
"source": "helmet"
|
| 1897 |
+
},
|
| 1898 |
+
"yi-1-5-9-b-32-k": {
|
| 1899 |
+
"display_name": "Yi-1.5-9B-32k",
|
| 1900 |
+
"ruler_per_ctx": null,
|
| 1901 |
+
"ruler_long_score": null,
|
| 1902 |
+
"helmet": {
|
| 1903 |
+
"overall": 2.1,
|
| 1904 |
+
"categories": {
|
| 1905 |
+
"Recall": 0.0,
|
| 1906 |
+
"RAG": 0.0,
|
| 1907 |
+
"Cite": 0.0,
|
| 1908 |
+
"Re-rank": 0.0,
|
| 1909 |
+
"ICL": 4.2,
|
| 1910 |
+
"LongQA": 7.1,
|
| 1911 |
+
"Summ": 3.2
|
| 1912 |
+
},
|
| 1913 |
+
"context_window_k": 32.0,
|
| 1914 |
+
"params_b": "9.0",
|
| 1915 |
+
"type": "♭"
|
| 1916 |
+
},
|
| 1917 |
+
"recipe_class": null,
|
| 1918 |
+
"params_b": "9.0",
|
| 1919 |
+
"native_context_k": null,
|
| 1920 |
+
"source": "helmet"
|
| 1921 |
+
},
|
| 1922 |
+
"phi-3-mini-128-k-instruct": {
|
| 1923 |
+
"display_name": "Phi-3-mini-128k-Inst",
|
| 1924 |
+
"ruler_per_ctx": null,
|
| 1925 |
+
"ruler_long_score": null,
|
| 1926 |
+
"helmet": {
|
| 1927 |
+
"overall": 33.6,
|
| 1928 |
+
"categories": {
|
| 1929 |
+
"Recall": 50.1,
|
| 1930 |
+
"RAG": 46.7,
|
| 1931 |
+
"Cite": 0.6,
|
| 1932 |
+
"Re-rank": 5.8,
|
| 1933 |
+
"ICL": 78.7,
|
| 1934 |
+
"LongQA": 29.9,
|
| 1935 |
+
"Summ": 23.7
|
| 1936 |
+
},
|
| 1937 |
+
"context_window_k": 128.0,
|
| 1938 |
+
"params_b": "4.0",
|
| 1939 |
+
"type": "♯"
|
| 1940 |
+
},
|
| 1941 |
+
"recipe_class": null,
|
| 1942 |
+
"params_b": "4.0",
|
| 1943 |
+
"native_context_k": null,
|
| 1944 |
+
"source": "helmet"
|
| 1945 |
+
},
|
| 1946 |
+
"phi-3-small-128-k-instruct": {
|
| 1947 |
+
"display_name": "Phi-3-small-128k-Inst",
|
| 1948 |
+
"ruler_per_ctx": null,
|
| 1949 |
+
"ruler_long_score": null,
|
| 1950 |
+
"helmet": {
|
| 1951 |
+
"overall": 24.9,
|
| 1952 |
+
"categories": {
|
| 1953 |
+
"Recall": 22.3,
|
| 1954 |
+
"RAG": 33.8,
|
| 1955 |
+
"Cite": 3.0,
|
| 1956 |
+
"Re-rank": 1.9,
|
| 1957 |
+
"ICL": 79.6,
|
| 1958 |
+
"LongQA": 27.5,
|
| 1959 |
+
"Summ": 6.6
|
| 1960 |
+
},
|
| 1961 |
+
"context_window_k": 128.0,
|
| 1962 |
+
"params_b": "7.0",
|
| 1963 |
+
"type": "♯"
|
| 1964 |
+
},
|
| 1965 |
+
"recipe_class": null,
|
| 1966 |
+
"params_b": "7.0",
|
| 1967 |
+
"native_context_k": null,
|
| 1968 |
+
"source": "helmet"
|
| 1969 |
+
},
|
| 1970 |
+
"phi-3-med-128-k-instruct": {
|
| 1971 |
+
"display_name": "Phi-3-med-128k-Inst",
|
| 1972 |
+
"ruler_per_ctx": null,
|
| 1973 |
+
"ruler_long_score": null,
|
| 1974 |
+
"helmet": {
|
| 1975 |
+
"overall": 29.1,
|
| 1976 |
+
"categories": {
|
| 1977 |
+
"Recall": 24.5,
|
| 1978 |
+
"RAG": 44.8,
|
| 1979 |
+
"Cite": 3.3,
|
| 1980 |
+
"Re-rank": 6.6,
|
| 1981 |
+
"ICL": 73.2,
|
| 1982 |
+
"LongQA": 26.9,
|
| 1983 |
+
"Summ": 24.8
|
| 1984 |
+
},
|
| 1985 |
+
"context_window_k": 128.0,
|
| 1986 |
+
"params_b": "14.0",
|
| 1987 |
+
"type": "♯"
|
| 1988 |
+
},
|
| 1989 |
+
"recipe_class": null,
|
| 1990 |
+
"params_b": "14.0",
|
| 1991 |
+
"native_context_k": null,
|
| 1992 |
+
"source": "helmet"
|
| 1993 |
+
},
|
| 1994 |
+
"phi-3-5-mini-instruct": {
|
| 1995 |
+
"display_name": "Phi-3.5-mini-Inst",
|
| 1996 |
+
"ruler_per_ctx": null,
|
| 1997 |
+
"ruler_long_score": null,
|
| 1998 |
+
"helmet": {
|
| 1999 |
+
"overall": 33.6,
|
| 2000 |
+
"categories": {
|
| 2001 |
+
"Recall": 48.8,
|
| 2002 |
+
"RAG": 43.1,
|
| 2003 |
+
"Cite": 1.6,
|
| 2004 |
+
"Re-rank": 7.8,
|
| 2005 |
+
"ICL": 79.5,
|
| 2006 |
+
"LongQA": 28.5,
|
| 2007 |
+
"Summ": 26.3
|
| 2008 |
+
},
|
| 2009 |
+
"context_window_k": 128.0,
|
| 2010 |
+
"params_b": "4.0",
|
| 2011 |
+
"type": "♯"
|
| 2012 |
+
},
|
| 2013 |
+
"recipe_class": null,
|
| 2014 |
+
"params_b": "4.0",
|
| 2015 |
+
"native_context_k": null,
|
| 2016 |
+
"source": "helmet"
|
| 2017 |
+
},
|
| 2018 |
+
"qwen-2-7-b": {
|
| 2019 |
+
"display_name": "Qwen2-7B",
|
| 2020 |
+
"ruler_per_ctx": null,
|
| 2021 |
+
"ruler_long_score": null,
|
| 2022 |
+
"helmet": {
|
| 2023 |
+
"overall": 30.0,
|
| 2024 |
+
"categories": {
|
| 2025 |
+
"Recall": 38.2,
|
| 2026 |
+
"RAG": 45.0,
|
| 2027 |
+
"Cite": 2.3,
|
| 2028 |
+
"Re-rank": 3.6,
|
| 2029 |
+
"ICL": 77.5,
|
| 2030 |
+
"LongQA": 36.8,
|
| 2031 |
+
"Summ": 6.8
|
| 2032 |
+
},
|
| 2033 |
+
"context_window_k": 32.0,
|
| 2034 |
+
"params_b": "7.0",
|
| 2035 |
+
"type": "♭"
|
| 2036 |
+
},
|
| 2037 |
+
"recipe_class": null,
|
| 2038 |
+
"params_b": "7.0",
|
| 2039 |
+
"native_context_k": null,
|
| 2040 |
+
"source": "helmet"
|
| 2041 |
+
},
|
| 2042 |
+
"qwen-2-7-b-instruct": {
|
| 2043 |
+
"display_name": "Qwen2-7B-Inst",
|
| 2044 |
+
"ruler_per_ctx": null,
|
| 2045 |
+
"ruler_long_score": null,
|
| 2046 |
+
"helmet": {
|
| 2047 |
+
"overall": 31.2,
|
| 2048 |
+
"categories": {
|
| 2049 |
+
"Recall": 36.8,
|
| 2050 |
+
"RAG": 47.4,
|
| 2051 |
+
"Cite": 2.4,
|
| 2052 |
+
"Re-rank": 5.9,
|
| 2053 |
+
"ICL": 66.2,
|
| 2054 |
+
"LongQA": 31.9,
|
| 2055 |
+
"Summ": 27.6
|
| 2056 |
+
},
|
| 2057 |
+
"context_window_k": 32.0,
|
| 2058 |
+
"params_b": "7.0",
|
| 2059 |
+
"type": "♯"
|
| 2060 |
+
},
|
| 2061 |
+
"recipe_class": null,
|
| 2062 |
+
"params_b": "7.0",
|
| 2063 |
+
"native_context_k": null,
|
| 2064 |
+
"source": "helmet"
|
| 2065 |
+
},
|
| 2066 |
+
"qwen-2-57-b": {
|
| 2067 |
+
"display_name": "Qwen2-57B",
|
| 2068 |
+
"ruler_per_ctx": null,
|
| 2069 |
+
"ruler_long_score": null,
|
| 2070 |
+
"helmet": {
|
| 2071 |
+
"overall": 11.0,
|
| 2072 |
+
"categories": {
|
| 2073 |
+
"Recall": 1.8,
|
| 2074 |
+
"RAG": 10.6,
|
| 2075 |
+
"Cite": 1.1,
|
| 2076 |
+
"Re-rank": 0.0,
|
| 2077 |
+
"ICL": 43.3,
|
| 2078 |
+
"LongQA": 15.7,
|
| 2079 |
+
"Summ": 4.9
|
| 2080 |
+
},
|
| 2081 |
+
"context_window_k": 32.0,
|
| 2082 |
+
"params_b": "14.0",
|
| 2083 |
+
"type": "♭"
|
| 2084 |
+
},
|
| 2085 |
+
"recipe_class": null,
|
| 2086 |
+
"params_b": "14.0",
|
| 2087 |
+
"native_context_k": null,
|
| 2088 |
+
"source": "helmet"
|
| 2089 |
+
},
|
| 2090 |
+
"qwen-2-57-b-instruct": {
|
| 2091 |
+
"display_name": "Qwen2-57B-Inst",
|
| 2092 |
+
"ruler_per_ctx": null,
|
| 2093 |
+
"ruler_long_score": null,
|
| 2094 |
+
"helmet": {
|
| 2095 |
+
"overall": 13.4,
|
| 2096 |
+
"categories": {
|
| 2097 |
+
"Recall": 5.9,
|
| 2098 |
+
"RAG": 12.7,
|
| 2099 |
+
"Cite": 1.1,
|
| 2100 |
+
"Re-rank": 0.0,
|
| 2101 |
+
"ICL": 40.6,
|
| 2102 |
+
"LongQA": 22.9,
|
| 2103 |
+
"Summ": 10.5
|
| 2104 |
+
},
|
| 2105 |
+
"context_window_k": 32.0,
|
| 2106 |
+
"params_b": "14.0",
|
| 2107 |
+
"type": "♯"
|
| 2108 |
+
},
|
| 2109 |
+
"recipe_class": null,
|
| 2110 |
+
"params_b": "14.0",
|
| 2111 |
+
"native_context_k": null,
|
| 2112 |
+
"source": "helmet"
|
| 2113 |
+
},
|
| 2114 |
+
"qwen-2-5-1-5-b": {
|
| 2115 |
+
"display_name": "Qwen2.5-1.5B",
|
| 2116 |
+
"ruler_per_ctx": null,
|
| 2117 |
+
"ruler_long_score": null,
|
| 2118 |
+
"helmet": {
|
| 2119 |
+
"overall": 12.9,
|
| 2120 |
+
"categories": {
|
| 2121 |
+
"Recall": 3.5,
|
| 2122 |
+
"RAG": 22.3,
|
| 2123 |
+
"Cite": 0.6,
|
| 2124 |
+
"Re-rank": 0.0,
|
| 2125 |
+
"ICL": 43.7,
|
| 2126 |
+
"LongQA": 18.2,
|
| 2127 |
+
"Summ": 2.1
|
| 2128 |
+
},
|
| 2129 |
+
"context_window_k": 32.0,
|
| 2130 |
+
"params_b": "1.5",
|
| 2131 |
+
"type": "♭"
|
| 2132 |
+
},
|
| 2133 |
+
"recipe_class": null,
|
| 2134 |
+
"params_b": "1.5",
|
| 2135 |
+
"native_context_k": null,
|
| 2136 |
+
"source": "helmet"
|
| 2137 |
+
},
|
| 2138 |
+
"qwen-2-5-1-5-b-instruct": {
|
| 2139 |
+
"display_name": "Qwen2.5-1.5B-Inst",
|
| 2140 |
+
"ruler_per_ctx": null,
|
| 2141 |
+
"ruler_long_score": null,
|
| 2142 |
+
"helmet": {
|
| 2143 |
+
"overall": 13.2,
|
| 2144 |
+
"categories": {
|
| 2145 |
+
"Recall": 7.8,
|
| 2146 |
+
"RAG": 25.0,
|
| 2147 |
+
"Cite": 1.7,
|
| 2148 |
+
"Re-rank": 0.3,
|
| 2149 |
+
"ICL": 38.4,
|
| 2150 |
+
"LongQA": 7.8,
|
| 2151 |
+
"Summ": 11.3
|
| 2152 |
+
},
|
| 2153 |
+
"context_window_k": 32.0,
|
| 2154 |
+
"params_b": "1.5",
|
| 2155 |
+
"type": "♯"
|
| 2156 |
+
},
|
| 2157 |
+
"recipe_class": null,
|
| 2158 |
+
"params_b": "1.5",
|
| 2159 |
+
"native_context_k": null,
|
| 2160 |
+
"source": "helmet"
|
| 2161 |
+
},
|
| 2162 |
+
"qwen-2-5-3-b": {
|
| 2163 |
+
"display_name": "Qwen2.5-3B",
|
| 2164 |
+
"ruler_per_ctx": null,
|
| 2165 |
+
"ruler_long_score": null,
|
| 2166 |
+
"helmet": {
|
| 2167 |
+
"overall": 12.3,
|
| 2168 |
+
"categories": {
|
| 2169 |
+
"Recall": 7.5,
|
| 2170 |
+
"RAG": 21.1,
|
| 2171 |
+
"Cite": 1.0,
|
| 2172 |
+
"Re-rank": 2.2,
|
| 2173 |
+
"ICL": 32.0,
|
| 2174 |
+
"LongQA": 20.5,
|
| 2175 |
+
"Summ": 1.8
|
| 2176 |
+
},
|
| 2177 |
+
"context_window_k": 32.0,
|
| 2178 |
+
"params_b": "3.0",
|
| 2179 |
+
"type": "♭"
|
| 2180 |
+
},
|
| 2181 |
+
"recipe_class": null,
|
| 2182 |
+
"params_b": "3.0",
|
| 2183 |
+
"native_context_k": null,
|
| 2184 |
+
"source": "helmet"
|
| 2185 |
+
},
|
| 2186 |
+
"qwen-2-5-3-b-instruct": {
|
| 2187 |
+
"display_name": "Qwen2.5-3B-Inst",
|
| 2188 |
+
"ruler_per_ctx": null,
|
| 2189 |
+
"ruler_long_score": null,
|
| 2190 |
+
"helmet": {
|
| 2191 |
+
"overall": 18.1,
|
| 2192 |
+
"categories": {
|
| 2193 |
+
"Recall": 10.4,
|
| 2194 |
+
"RAG": 22.4,
|
| 2195 |
+
"Cite": 1.9,
|
| 2196 |
+
"Re-rank": 3.7,
|
| 2197 |
+
"ICL": 49.4,
|
| 2198 |
+
"LongQA": 21.3,
|
| 2199 |
+
"Summ": 17.5
|
| 2200 |
+
},
|
| 2201 |
+
"context_window_k": 32.0,
|
| 2202 |
+
"params_b": "3.0",
|
| 2203 |
+
"type": "♯"
|
| 2204 |
+
},
|
| 2205 |
+
"recipe_class": null,
|
| 2206 |
+
"params_b": "3.0",
|
| 2207 |
+
"native_context_k": null,
|
| 2208 |
+
"source": "helmet"
|
| 2209 |
+
},
|
| 2210 |
+
"qwen-2-5-7-b": {
|
| 2211 |
+
"display_name": "Qwen2.5-7B",
|
| 2212 |
+
"ruler_per_ctx": null,
|
| 2213 |
+
"ruler_long_score": null,
|
| 2214 |
+
"helmet": {
|
| 2215 |
+
"overall": 20.9,
|
| 2216 |
+
"categories": {
|
| 2217 |
+
"Recall": 13.9,
|
| 2218 |
+
"RAG": 27.0,
|
| 2219 |
+
"Cite": 1.3,
|
| 2220 |
+
"Re-rank": 0.1,
|
| 2221 |
+
"ICL": 71.5,
|
| 2222 |
+
"LongQA": 25.9,
|
| 2223 |
+
"Summ": 6.8
|
| 2224 |
+
},
|
| 2225 |
+
"context_window_k": 128.0,
|
| 2226 |
+
"params_b": "7.0",
|
| 2227 |
+
"type": "♭"
|
| 2228 |
+
},
|
| 2229 |
+
"recipe_class": null,
|
| 2230 |
+
"params_b": "7.0",
|
| 2231 |
+
"native_context_k": null,
|
| 2232 |
+
"source": "helmet"
|
| 2233 |
+
},
|
| 2234 |
+
"qwen-2-5-72-b-instruct": {
|
| 2235 |
+
"display_name": "Qwen2.5-72B-Inst",
|
| 2236 |
+
"ruler_per_ctx": null,
|
| 2237 |
+
"ruler_long_score": null,
|
| 2238 |
+
"helmet": {
|
| 2239 |
+
"overall": 38.2,
|
| 2240 |
+
"categories": {
|
| 2241 |
+
"Recall": 38.4,
|
| 2242 |
+
"RAG": 43.0,
|
| 2243 |
+
"Cite": 8.0,
|
| 2244 |
+
"Re-rank": 24.5,
|
| 2245 |
+
"ICL": 83.2,
|
| 2246 |
+
"LongQA": 38.9,
|
| 2247 |
+
"Summ": 31.4
|
| 2248 |
+
},
|
| 2249 |
+
"context_window_k": 128.0,
|
| 2250 |
+
"params_b": "72.0",
|
| 2251 |
+
"type": "♯"
|
| 2252 |
+
},
|
| 2253 |
+
"recipe_class": null,
|
| 2254 |
+
"params_b": "72.0",
|
| 2255 |
+
"native_context_k": null,
|
| 2256 |
+
"source": "helmet"
|
| 2257 |
+
},
|
| 2258 |
+
"prolong": {
|
| 2259 |
+
"display_name": "ProLong",
|
| 2260 |
+
"ruler_per_ctx": null,
|
| 2261 |
+
"ruler_long_score": null,
|
| 2262 |
+
"helmet": {
|
| 2263 |
+
"overall": 49.4,
|
| 2264 |
+
"categories": {
|
| 2265 |
+
"Recall": 98.8,
|
| 2266 |
+
"RAG": 63.2,
|
| 2267 |
+
"Cite": 1.4,
|
| 2268 |
+
"Re-rank": 22.5,
|
| 2269 |
+
"ICL": 86.5,
|
| 2270 |
+
"LongQA": 43.9,
|
| 2271 |
+
"Summ": 29.2
|
| 2272 |
+
},
|
| 2273 |
+
"context_window_k": 512.0,
|
| 2274 |
+
"params_b": "8.0",
|
| 2275 |
+
"type": "♯"
|
| 2276 |
+
},
|
| 2277 |
+
"recipe_class": null,
|
| 2278 |
+
"params_b": "8.0",
|
| 2279 |
+
"native_context_k": null,
|
| 2280 |
+
"source": "helmet"
|
| 2281 |
+
},
|
| 2282 |
+
"gemma-3-12-b": {
|
| 2283 |
+
"display_name": "Gemma-3-12B",
|
| 2284 |
+
"ruler_per_ctx": null,
|
| 2285 |
+
"ruler_long_score": null,
|
| 2286 |
+
"helmet": {
|
| 2287 |
+
"overall": 40.6,
|
| 2288 |
+
"categories": {
|
| 2289 |
+
"Recall": 76.6,
|
| 2290 |
+
"RAG": 59.3,
|
| 2291 |
+
"Cite": 2.0,
|
| 2292 |
+
"Re-rank": 16.0,
|
| 2293 |
+
"ICL": 88.7,
|
| 2294 |
+
"LongQA": 41.4,
|
| 2295 |
+
"Summ": 0.2
|
| 2296 |
+
},
|
| 2297 |
+
"context_window_k": 128.0,
|
| 2298 |
+
"params_b": "12.0",
|
| 2299 |
+
"type": "♭"
|
| 2300 |
+
},
|
| 2301 |
+
"recipe_class": null,
|
| 2302 |
+
"params_b": "12.0",
|
| 2303 |
+
"native_context_k": null,
|
| 2304 |
+
"source": "helmet"
|
| 2305 |
+
},
|
| 2306 |
+
"gemma-3-12-b-instruct": {
|
| 2307 |
+
"display_name": "Gemma-3-12B-Inst",
|
| 2308 |
+
"ruler_per_ctx": null,
|
| 2309 |
+
"ruler_long_score": null,
|
| 2310 |
+
"helmet": {
|
| 2311 |
+
"overall": 36.0,
|
| 2312 |
+
"categories": {
|
| 2313 |
+
"Recall": 25.9,
|
| 2314 |
+
"RAG": 52.1,
|
| 2315 |
+
"Cite": 1.3,
|
| 2316 |
+
"Re-rank": 21.3,
|
| 2317 |
+
"ICL": 75.4,
|
| 2318 |
+
"LongQA": 43.0,
|
| 2319 |
+
"Summ": 33.1
|
| 2320 |
+
},
|
| 2321 |
+
"context_window_k": 128.0,
|
| 2322 |
+
"params_b": "12.0",
|
| 2323 |
+
"type": "♯"
|
| 2324 |
+
},
|
| 2325 |
+
"recipe_class": null,
|
| 2326 |
+
"params_b": "12.0",
|
| 2327 |
+
"native_context_k": null,
|
| 2328 |
+
"source": "helmet"
|
| 2329 |
+
},
|
| 2330 |
+
"gemma-3-27-b": {
|
| 2331 |
+
"display_name": "Gemma-3-27B",
|
| 2332 |
+
"ruler_per_ctx": null,
|
| 2333 |
+
"ruler_long_score": null,
|
| 2334 |
+
"helmet": {
|
| 2335 |
+
"overall": 44.5,
|
| 2336 |
+
"categories": {
|
| 2337 |
+
"Recall": 82.6,
|
| 2338 |
+
"RAG": 64.1,
|
| 2339 |
+
"Cite": 2.8,
|
| 2340 |
+
"Re-rank": 30.4,
|
| 2341 |
+
"ICL": 86.3,
|
| 2342 |
+
"LongQA": 44.6,
|
| 2343 |
+
"Summ": 0.7
|
| 2344 |
+
},
|
| 2345 |
+
"context_window_k": 128.0,
|
| 2346 |
+
"params_b": "27.0",
|
| 2347 |
+
"type": "♭"
|
| 2348 |
+
},
|
| 2349 |
+
"recipe_class": null,
|
| 2350 |
+
"params_b": "27.0",
|
| 2351 |
+
"native_context_k": null,
|
| 2352 |
+
"source": "helmet"
|
| 2353 |
+
},
|
| 2354 |
+
"gemma-3-27-b-instruct": {
|
| 2355 |
+
"display_name": "Gemma-3-27B-Inst",
|
| 2356 |
+
"ruler_per_ctx": null,
|
| 2357 |
+
"ruler_long_score": null,
|
| 2358 |
+
"helmet": {
|
| 2359 |
+
"overall": 41.9,
|
| 2360 |
+
"categories": {
|
| 2361 |
+
"Recall": 45.1,
|
| 2362 |
+
"RAG": 56.6,
|
| 2363 |
+
"Cite": 1.6,
|
| 2364 |
+
"Re-rank": 30.0,
|
| 2365 |
+
"ICL": 77.4,
|
| 2366 |
+
"LongQA": 46.8,
|
| 2367 |
+
"Summ": 35.9
|
| 2368 |
+
},
|
| 2369 |
+
"context_window_k": 128.0,
|
| 2370 |
+
"params_b": "27.0",
|
| 2371 |
+
"type": "♯"
|
| 2372 |
+
},
|
| 2373 |
+
"recipe_class": null,
|
| 2374 |
+
"params_b": "27.0",
|
| 2375 |
+
"native_context_k": null,
|
| 2376 |
+
"source": "helmet"
|
| 2377 |
+
},
|
| 2378 |
+
"jamba-v-0-1": {
|
| 2379 |
+
"display_name": "Jamba-v0.1",
|
| 2380 |
+
"ruler_per_ctx": null,
|
| 2381 |
+
"ruler_long_score": null,
|
| 2382 |
+
"helmet": {
|
| 2383 |
+
"overall": 41.5,
|
| 2384 |
+
"categories": {
|
| 2385 |
+
"Recall": 80.3,
|
| 2386 |
+
"RAG": 60.3,
|
| 2387 |
+
"Cite": 6.0,
|
| 2388 |
+
"Re-rank": 9.3,
|
| 2389 |
+
"ICL": 88.7,
|
| 2390 |
+
"LongQA": 41.7,
|
| 2391 |
+
"Summ": 4.0
|
| 2392 |
+
},
|
| 2393 |
+
"context_window_k": 256.0,
|
| 2394 |
+
"params_b": "12.0",
|
| 2395 |
+
"type": "♭"
|
| 2396 |
+
},
|
| 2397 |
+
"recipe_class": null,
|
| 2398 |
+
"params_b": "12.0",
|
| 2399 |
+
"native_context_k": null,
|
| 2400 |
+
"source": "helmet"
|
| 2401 |
+
}
|
| 2402 |
+
},
|
| 2403 |
+
"aliases": {
|
| 2404 |
+
"qwen-2-5-7-b-instruct": [
|
| 2405 |
+
"qwen2.5-7b-instruct",
|
| 2406 |
+
"Qwen2.5-7B-Inst",
|
| 2407 |
+
"Qwen2.5-7B-Inst",
|
| 2408 |
+
"Qwen2.5-7B-Inst",
|
| 2409 |
+
"Qwen2.5-7B-Inst",
|
| 2410 |
+
"Qwen2.5-7B-Inst"
|
| 2411 |
+
],
|
| 2412 |
+
"qwen-2-5-14-b-instruct": [
|
| 2413 |
+
"qwen2.5-14b-instruct"
|
| 2414 |
+
],
|
| 2415 |
+
"qwen-2-5-32-b-instruct": [
|
| 2416 |
+
"qwen2.5-32b-instruct"
|
| 2417 |
+
],
|
| 2418 |
+
"mistral-7-b-v-0-2": [
|
| 2419 |
+
"mistral-7b-v0.2"
|
| 2420 |
+
],
|
| 2421 |
+
"mixtral-8-x-7-b": [
|
| 2422 |
+
"mixtral-8x7b"
|
| 2423 |
+
],
|
| 2424 |
+
"mixtral-8-x-22-b-instruct": [
|
| 2425 |
+
"mixtral-8x22b-instruct"
|
| 2426 |
+
],
|
| 2427 |
+
"qwen-2-72-b": [
|
| 2428 |
+
"qwen2-72b"
|
| 2429 |
+
],
|
| 2430 |
+
"dbrx": [
|
| 2431 |
+
"dbrx"
|
| 2432 |
+
],
|
| 2433 |
+
"qwen-1-5-72-b": [
|
| 2434 |
+
"qwen1.5-72b"
|
| 2435 |
+
],
|
| 2436 |
+
"mistral-large-2407": [
|
| 2437 |
+
"mistral-large-2407"
|
| 2438 |
+
],
|
| 2439 |
+
"yi-34-b-200-k": [
|
| 2440 |
+
"yi-34b-200k",
|
| 2441 |
+
"Yi-34B-200k",
|
| 2442 |
+
"Yi-34B-200k",
|
| 2443 |
+
"Yi-34B-200k",
|
| 2444 |
+
"Yi-34B-200k",
|
| 2445 |
+
"Yi-34B-200k"
|
| 2446 |
+
],
|
| 2447 |
+
"command-r-35-b": [
|
| 2448 |
+
"command-r-35b"
|
| 2449 |
+
],
|
| 2450 |
+
"command-r-0824-32-b": [
|
| 2451 |
+
"command-r-0824-32b"
|
| 2452 |
+
],
|
| 2453 |
+
"command-r-plus-104-b": [
|
| 2454 |
+
"command-r-plus-104b"
|
| 2455 |
+
],
|
| 2456 |
+
"command-r-plus-0824-104-b": [
|
| 2457 |
+
"command-r-plus-0824-104b"
|
| 2458 |
+
],
|
| 2459 |
+
"mistral-large-2411-123-b": [
|
| 2460 |
+
"mistral-large-2411-123b"
|
| 2461 |
+
],
|
| 2462 |
+
"chatglm-3-6-b-128-k": [
|
| 2463 |
+
"chatglm3-6b-128k"
|
| 2464 |
+
],
|
| 2465 |
+
"lwm-7-b": [
|
| 2466 |
+
"lwm-7b"
|
| 2467 |
+
],
|
| 2468 |
+
"glm-4-9-b-1-m": [
|
| 2469 |
+
"glm4-9b-1m"
|
| 2470 |
+
],
|
| 2471 |
+
"prolong-8-b-512-k": [
|
| 2472 |
+
"prolong-8b-512k"
|
| 2473 |
+
],
|
| 2474 |
+
"megabeam-mistral-7-b-512-k": [
|
| 2475 |
+
"megabeam-mistral-7b-512k"
|
| 2476 |
+
],
|
| 2477 |
+
"internlm-2-5-7-b-1-m": [
|
| 2478 |
+
"internlm2.5-7b-1m"
|
| 2479 |
+
],
|
| 2480 |
+
"gradientai-llama-3-70-b-1-m": [
|
| 2481 |
+
"gradientai-llama3-70b-1m"
|
| 2482 |
+
],
|
| 2483 |
+
"gradientai-llama-3-8-b-1-m": [
|
| 2484 |
+
"gradientai-llama3-8b-1m"
|
| 2485 |
+
],
|
| 2486 |
+
"longchat-7-b-32-k": [
|
| 2487 |
+
"longchat-7b-32k"
|
| 2488 |
+
],
|
| 2489 |
+
"longalpaca-13-b-32-k": [
|
| 2490 |
+
"longalpaca-13b-32k"
|
| 2491 |
+
],
|
| 2492 |
+
"llama-3-1-70-b-instruct": [
|
| 2493 |
+
"llama3.1-70b-instruct",
|
| 2494 |
+
"Llama-3.1-70B-Inst",
|
| 2495 |
+
"Llama-3.1-70B-Inst",
|
| 2496 |
+
"Llama-3.1-70B-Inst",
|
| 2497 |
+
"Llama-3.1-70B-Inst",
|
| 2498 |
+
"Llama-3.1-70B-Inst"
|
| 2499 |
+
],
|
| 2500 |
+
"llama-3-1-8-b-instruct": [
|
| 2501 |
+
"llama3.1-8b-instruct",
|
| 2502 |
+
"Llama-3.1-8B-Inst",
|
| 2503 |
+
"Llama-3.1-8B-Inst",
|
| 2504 |
+
"Llama-3.1-8B-Inst",
|
| 2505 |
+
"Llama-3.1-8B-Inst",
|
| 2506 |
+
"Llama-3.1-8B-Inst"
|
| 2507 |
+
],
|
| 2508 |
+
"mistral-nemo-12-b": [
|
| 2509 |
+
"mistral-nemo-12b"
|
| 2510 |
+
],
|
| 2511 |
+
"qwen-2-5-7-b-instruct-1-m": [
|
| 2512 |
+
"qwen2.5-7b-instruct-1m"
|
| 2513 |
+
],
|
| 2514 |
+
"qwen-2-5-14-b-instruct-1-m": [
|
| 2515 |
+
"qwen2.5-14b-instruct-1m"
|
| 2516 |
+
],
|
| 2517 |
+
"jamba-1-5-large": [
|
| 2518 |
+
"jamba-1.5-large"
|
| 2519 |
+
],
|
| 2520 |
+
"jamba-1-5-mini": [
|
| 2521 |
+
"jamba-1.5-mini",
|
| 2522 |
+
"Jamba-1.5-Mini",
|
| 2523 |
+
"Jamba-1.5-Mini",
|
| 2524 |
+
"Jamba-1.5-Mini",
|
| 2525 |
+
"Jamba-1.5-Mini",
|
| 2526 |
+
"Jamba-1.5-Mini"
|
| 2527 |
+
],
|
| 2528 |
+
"phi-3-mini-3-8-b": [
|
| 2529 |
+
"phi3-mini-3.8b"
|
| 2530 |
+
],
|
| 2531 |
+
"phi-3-medium-14-b": [
|
| 2532 |
+
"phi3-medium-14b"
|
| 2533 |
+
],
|
| 2534 |
+
"gpt-4": [
|
| 2535 |
+
"GPT-4",
|
| 2536 |
+
"GPT-4",
|
| 2537 |
+
"GPT-4",
|
| 2538 |
+
"GPT-4",
|
| 2539 |
+
"GPT-4"
|
| 2540 |
+
],
|
| 2541 |
+
"gpt-4-o-mini": [
|
| 2542 |
+
"GPT-4o-mini",
|
| 2543 |
+
"GPT-4o-mini",
|
| 2544 |
+
"GPT-4o-mini",
|
| 2545 |
+
"GPT-4o-mini",
|
| 2546 |
+
"GPT-4o-mini"
|
| 2547 |
+
],
|
| 2548 |
+
"gpt-4-o-05": [
|
| 2549 |
+
"GPT-4o-05",
|
| 2550 |
+
"GPT-4o-05",
|
| 2551 |
+
"GPT-4o-05",
|
| 2552 |
+
"GPT-4o-05",
|
| 2553 |
+
"GPT-4o-05"
|
| 2554 |
+
],
|
| 2555 |
+
"gpt-4-o-08": [
|
| 2556 |
+
"GPT-4o-08",
|
| 2557 |
+
"GPT-4o-08",
|
| 2558 |
+
"GPT-4o-08",
|
| 2559 |
+
"GPT-4o-08",
|
| 2560 |
+
"GPT-4o-08"
|
| 2561 |
+
],
|
| 2562 |
+
"claude-3-5-sonnet": [
|
| 2563 |
+
"Claude-3.5-Sonnet",
|
| 2564 |
+
"Claude-3.5-Sonnet",
|
| 2565 |
+
"Claude-3.5-Sonnet",
|
| 2566 |
+
"Claude-3.5-Sonnet",
|
| 2567 |
+
"Claude-3.5-Sonnet"
|
| 2568 |
+
],
|
| 2569 |
+
"gemini-1-5-flash": [
|
| 2570 |
+
"Gemini-1.5-Flash",
|
| 2571 |
+
"Gemini-1.5-Flash",
|
| 2572 |
+
"Gemini-1.5-Flash",
|
| 2573 |
+
"Gemini-1.5-Flash",
|
| 2574 |
+
"Gemini-1.5-Flash"
|
| 2575 |
+
],
|
| 2576 |
+
"gemini-1-5-pro": [
|
| 2577 |
+
"Gemini-1.5-Pro",
|
| 2578 |
+
"Gemini-1.5-Pro",
|
| 2579 |
+
"Gemini-1.5-Pro",
|
| 2580 |
+
"Gemini-1.5-Pro",
|
| 2581 |
+
"Gemini-1.5-Pro"
|
| 2582 |
+
],
|
| 2583 |
+
"llama-2-7-b-32-k": [
|
| 2584 |
+
"Llama-2-7B-32k",
|
| 2585 |
+
"Llama-2-7B-32k",
|
| 2586 |
+
"Llama-2-7B-32k",
|
| 2587 |
+
"Llama-2-7B-32k",
|
| 2588 |
+
"Llama-2-7B-32k"
|
| 2589 |
+
],
|
| 2590 |
+
"llama-2-7-b-32-k-instruct": [
|
| 2591 |
+
"Llama-2-7B-32k-Inst",
|
| 2592 |
+
"Llama-2-7B-32k-Inst",
|
| 2593 |
+
"Llama-2-7B-32k-Inst",
|
| 2594 |
+
"Llama-2-7B-32k-Inst",
|
| 2595 |
+
"Llama-2-7B-32k-Inst"
|
| 2596 |
+
],
|
| 2597 |
+
"llama-7-b-80-k": [
|
| 2598 |
+
"Llama-7B-80k",
|
| 2599 |
+
"Llama-7B-80k",
|
| 2600 |
+
"Llama-7B-80k",
|
| 2601 |
+
"Llama-7B-80k",
|
| 2602 |
+
"Llama-7B-80k"
|
| 2603 |
+
],
|
| 2604 |
+
"yarn-llama-2-7-b-64-k": [
|
| 2605 |
+
"Yarn-Llama-2-7B-64k",
|
| 2606 |
+
"Yarn-Llama-2-7B-64k",
|
| 2607 |
+
"Yarn-Llama-2-7B-64k",
|
| 2608 |
+
"Yarn-Llama-2-7B-64k",
|
| 2609 |
+
"Yarn-Llama-2-7B-64k"
|
| 2610 |
+
],
|
| 2611 |
+
"yarn-llama-2-7-b-128-k": [
|
| 2612 |
+
"Yarn-Llama-2-7B-128k",
|
| 2613 |
+
"Yarn-Llama-2-7B-128k",
|
| 2614 |
+
"Yarn-Llama-2-7B-128k",
|
| 2615 |
+
"Yarn-Llama-2-7B-128k",
|
| 2616 |
+
"Yarn-Llama-2-7B-128k"
|
| 2617 |
+
],
|
| 2618 |
+
"llama-3-8-b": [
|
| 2619 |
+
"Llama-3-8B",
|
| 2620 |
+
"Llama-3-8B",
|
| 2621 |
+
"Llama-3-8B",
|
| 2622 |
+
"Llama-3-8B",
|
| 2623 |
+
"Llama-3-8B"
|
| 2624 |
+
],
|
| 2625 |
+
"llama-3-8-b-instruct": [
|
| 2626 |
+
"Llama-3-8B-Inst",
|
| 2627 |
+
"Llama-3-8B-Inst",
|
| 2628 |
+
"Llama-3-8B-Inst",
|
| 2629 |
+
"Llama-3-8B-Inst",
|
| 2630 |
+
"Llama-3-8B-Inst"
|
| 2631 |
+
],
|
| 2632 |
+
"llama-3-8-b-θ": [
|
| 2633 |
+
"Llama-3-8B-θ",
|
| 2634 |
+
"Llama-3-8B-θ",
|
| 2635 |
+
"Llama-3-8B-θ",
|
| 2636 |
+
"Llama-3-8B-θ",
|
| 2637 |
+
"Llama-3-8B-θ"
|
| 2638 |
+
],
|
| 2639 |
+
"llama-3-8-b-instruct-θ": [
|
| 2640 |
+
"Llama-3-8B-Inst-θ",
|
| 2641 |
+
"Llama-3-8B-Inst-θ",
|
| 2642 |
+
"Llama-3-8B-Inst-θ",
|
| 2643 |
+
"Llama-3-8B-Inst-θ",
|
| 2644 |
+
"Llama-3-8B-Inst-θ"
|
| 2645 |
+
],
|
| 2646 |
+
"llama-3-70-b-θ": [
|
| 2647 |
+
"Llama-3-70B-θ",
|
| 2648 |
+
"Llama-3-70B-θ",
|
| 2649 |
+
"Llama-3-70B-θ",
|
| 2650 |
+
"Llama-3-70B-θ",
|
| 2651 |
+
"Llama-3-70B-θ"
|
| 2652 |
+
],
|
| 2653 |
+
"llama-3-70-b-instruct-θ": [
|
| 2654 |
+
"Llama-3-70B-Inst-θ",
|
| 2655 |
+
"Llama-3-70B-Inst-θ",
|
| 2656 |
+
"Llama-3-70B-Inst-θ",
|
| 2657 |
+
"Llama-3-70B-Inst-θ",
|
| 2658 |
+
"Llama-3-70B-Inst-θ"
|
| 2659 |
+
],
|
| 2660 |
+
"llama-3-1-8-b": [
|
| 2661 |
+
"Llama-3.1-8B",
|
| 2662 |
+
"Llama-3.1-8B",
|
| 2663 |
+
"Llama-3.1-8B",
|
| 2664 |
+
"Llama-3.1-8B",
|
| 2665 |
+
"Llama-3.1-8B"
|
| 2666 |
+
],
|
| 2667 |
+
"llama-3-1-70-b": [
|
| 2668 |
+
"Llama-3.1-70B",
|
| 2669 |
+
"Llama-3.1-70B",
|
| 2670 |
+
"Llama-3.1-70B",
|
| 2671 |
+
"Llama-3.1-70B",
|
| 2672 |
+
"Llama-3.1-70B"
|
| 2673 |
+
],
|
| 2674 |
+
"llama-3-3-70-b-instruct": [
|
| 2675 |
+
"Llama-3.3-70B-Inst",
|
| 2676 |
+
"Llama-3.3-70B-Inst",
|
| 2677 |
+
"Llama-3.3-70B-Inst",
|
| 2678 |
+
"Llama-3.3-70B-Inst",
|
| 2679 |
+
"Llama-3.3-70B-Inst"
|
| 2680 |
+
],
|
| 2681 |
+
"llama-3-2-1-b": [
|
| 2682 |
+
"Llama-3.2-1B",
|
| 2683 |
+
"Llama-3.2-1B",
|
| 2684 |
+
"Llama-3.2-1B",
|
| 2685 |
+
"Llama-3.2-1B",
|
| 2686 |
+
"Llama-3.2-1B"
|
| 2687 |
+
],
|
| 2688 |
+
"llama-3-2-1-b-instruct": [
|
| 2689 |
+
"Llama-3.2-1B-Inst",
|
| 2690 |
+
"Llama-3.2-1B-Inst",
|
| 2691 |
+
"Llama-3.2-1B-Inst",
|
| 2692 |
+
"Llama-3.2-1B-Inst",
|
| 2693 |
+
"Llama-3.2-1B-Inst"
|
| 2694 |
+
],
|
| 2695 |
+
"llama-3-2-3-b": [
|
| 2696 |
+
"Llama-3.2-3B",
|
| 2697 |
+
"Llama-3.2-3B",
|
| 2698 |
+
"Llama-3.2-3B",
|
| 2699 |
+
"Llama-3.2-3B",
|
| 2700 |
+
"Llama-3.2-3B"
|
| 2701 |
+
],
|
| 2702 |
+
"llama-3-2-3-b-instruct": [
|
| 2703 |
+
"Llama-3.2-3B-Inst",
|
| 2704 |
+
"Llama-3.2-3B-Inst",
|
| 2705 |
+
"Llama-3.2-3B-Inst",
|
| 2706 |
+
"Llama-3.2-3B-Inst",
|
| 2707 |
+
"Llama-3.2-3B-Inst"
|
| 2708 |
+
],
|
| 2709 |
+
"llama-4-17-b": [
|
| 2710 |
+
"Llama-4-17B",
|
| 2711 |
+
"Llama-4-17B",
|
| 2712 |
+
"Llama-4-17B",
|
| 2713 |
+
"Llama-4-17B",
|
| 2714 |
+
"Llama-4-17B"
|
| 2715 |
+
],
|
| 2716 |
+
"llama-4-17-b-instruct": [
|
| 2717 |
+
"Llama-4-17B-Inst",
|
| 2718 |
+
"Llama-4-17B-Inst",
|
| 2719 |
+
"Llama-4-17B-Inst",
|
| 2720 |
+
"Llama-4-17B-Inst",
|
| 2721 |
+
"Llama-4-17B-Inst"
|
| 2722 |
+
],
|
| 2723 |
+
"mistral-7-b-instruct-v-0-1": [
|
| 2724 |
+
"Mistral-7B-Inst-v0.1",
|
| 2725 |
+
"Mistral-7B-Inst-v0.1",
|
| 2726 |
+
"Mistral-7B-Inst-v0.1",
|
| 2727 |
+
"Mistral-7B-Inst-v0.1",
|
| 2728 |
+
"Mistral-7B-Inst-v0.1"
|
| 2729 |
+
],
|
| 2730 |
+
"mistral-7-b-instruct-v-0-2": [
|
| 2731 |
+
"Mistral-7B-Inst-v0.2",
|
| 2732 |
+
"Mistral-7B-Inst-v0.2",
|
| 2733 |
+
"Mistral-7B-Inst-v0.2",
|
| 2734 |
+
"Mistral-7B-Inst-v0.2",
|
| 2735 |
+
"Mistral-7B-Inst-v0.2"
|
| 2736 |
+
],
|
| 2737 |
+
"mistral-7-b-v-0-3": [
|
| 2738 |
+
"Mistral-7B-v0.3",
|
| 2739 |
+
"Mistral-7B-v0.3",
|
| 2740 |
+
"Mistral-7B-v0.3",
|
| 2741 |
+
"Mistral-7B-v0.3",
|
| 2742 |
+
"Mistral-7B-v0.3"
|
| 2743 |
+
],
|
| 2744 |
+
"mistral-7-b-instruct-v-0-3": [
|
| 2745 |
+
"Mistral-7B-Inst-v0.3",
|
| 2746 |
+
"Mistral-7B-Inst-v0.3",
|
| 2747 |
+
"Mistral-7B-Inst-v0.3",
|
| 2748 |
+
"Mistral-7B-Inst-v0.3",
|
| 2749 |
+
"Mistral-7B-Inst-v0.3"
|
| 2750 |
+
],
|
| 2751 |
+
"ministral-8-b-instruct": [
|
| 2752 |
+
"Ministral-8B-Inst",
|
| 2753 |
+
"Ministral-8B-Inst",
|
| 2754 |
+
"Ministral-8B-Inst",
|
| 2755 |
+
"Ministral-8B-Inst",
|
| 2756 |
+
"Ministral-8B-Inst"
|
| 2757 |
+
],
|
| 2758 |
+
"mistral-nemo": [
|
| 2759 |
+
"Mistral-Nemo",
|
| 2760 |
+
"Mistral-Nemo",
|
| 2761 |
+
"Mistral-Nemo",
|
| 2762 |
+
"Mistral-Nemo",
|
| 2763 |
+
"Mistral-Nemo"
|
| 2764 |
+
],
|
| 2765 |
+
"mistral-nemo-instruct": [
|
| 2766 |
+
"Mistral-Nemo-Inst",
|
| 2767 |
+
"Mistral-Nemo-Inst",
|
| 2768 |
+
"Mistral-Nemo-Inst",
|
| 2769 |
+
"Mistral-Nemo-Inst",
|
| 2770 |
+
"Mistral-Nemo-Inst"
|
| 2771 |
+
],
|
| 2772 |
+
"megabeam-mistral": [
|
| 2773 |
+
"MegaBeam-Mistral",
|
| 2774 |
+
"MegaBeam-Mistral",
|
| 2775 |
+
"MegaBeam-Mistral",
|
| 2776 |
+
"MegaBeam-Mistral",
|
| 2777 |
+
"MegaBeam-Mistral"
|
| 2778 |
+
],
|
| 2779 |
+
"yi-6-b-200-k": [
|
| 2780 |
+
"Yi-6B-200k",
|
| 2781 |
+
"Yi-6B-200k",
|
| 2782 |
+
"Yi-6B-200k",
|
| 2783 |
+
"Yi-6B-200k",
|
| 2784 |
+
"Yi-6B-200k"
|
| 2785 |
+
],
|
| 2786 |
+
"yi-9-b-200-k": [
|
| 2787 |
+
"Yi-9B-200k",
|
| 2788 |
+
"Yi-9B-200k",
|
| 2789 |
+
"Yi-9B-200k",
|
| 2790 |
+
"Yi-9B-200k",
|
| 2791 |
+
"Yi-9B-200k"
|
| 2792 |
+
],
|
| 2793 |
+
"yi-1-5-9-b-32-k": [
|
| 2794 |
+
"Yi-1.5-9B-32k",
|
| 2795 |
+
"Yi-1.5-9B-32k",
|
| 2796 |
+
"Yi-1.5-9B-32k",
|
| 2797 |
+
"Yi-1.5-9B-32k",
|
| 2798 |
+
"Yi-1.5-9B-32k"
|
| 2799 |
+
],
|
| 2800 |
+
"phi-3-mini-128-k-instruct": [
|
| 2801 |
+
"Phi-3-mini-128k-Inst",
|
| 2802 |
+
"Phi-3-mini-128k-Inst",
|
| 2803 |
+
"Phi-3-mini-128k-Inst",
|
| 2804 |
+
"Phi-3-mini-128k-Inst",
|
| 2805 |
+
"Phi-3-mini-128k-Inst"
|
| 2806 |
+
],
|
| 2807 |
+
"phi-3-small-128-k-instruct": [
|
| 2808 |
+
"Phi-3-small-128k-Inst",
|
| 2809 |
+
"Phi-3-small-128k-Inst",
|
| 2810 |
+
"Phi-3-small-128k-Inst",
|
| 2811 |
+
"Phi-3-small-128k-Inst",
|
| 2812 |
+
"Phi-3-small-128k-Inst"
|
| 2813 |
+
],
|
| 2814 |
+
"phi-3-med-128-k-instruct": [
|
| 2815 |
+
"Phi-3-med-128k-Inst",
|
| 2816 |
+
"Phi-3-med-128k-Inst",
|
| 2817 |
+
"Phi-3-med-128k-Inst",
|
| 2818 |
+
"Phi-3-med-128k-Inst",
|
| 2819 |
+
"Phi-3-med-128k-Inst"
|
| 2820 |
+
],
|
| 2821 |
+
"phi-3-5-mini-instruct": [
|
| 2822 |
+
"Phi-3.5-mini-Inst",
|
| 2823 |
+
"Phi-3.5-mini-Inst",
|
| 2824 |
+
"Phi-3.5-mini-Inst",
|
| 2825 |
+
"Phi-3.5-mini-Inst",
|
| 2826 |
+
"Phi-3.5-mini-Inst"
|
| 2827 |
+
],
|
| 2828 |
+
"qwen-2-7-b": [
|
| 2829 |
+
"Qwen2-7B",
|
| 2830 |
+
"Qwen2-7B",
|
| 2831 |
+
"Qwen2-7B",
|
| 2832 |
+
"Qwen2-7B",
|
| 2833 |
+
"Qwen2-7B"
|
| 2834 |
+
],
|
| 2835 |
+
"qwen-2-7-b-instruct": [
|
| 2836 |
+
"Qwen2-7B-Inst",
|
| 2837 |
+
"Qwen2-7B-Inst",
|
| 2838 |
+
"Qwen2-7B-Inst",
|
| 2839 |
+
"Qwen2-7B-Inst",
|
| 2840 |
+
"Qwen2-7B-Inst"
|
| 2841 |
+
],
|
| 2842 |
+
"qwen-2-57-b": [
|
| 2843 |
+
"Qwen2-57B",
|
| 2844 |
+
"Qwen2-57B",
|
| 2845 |
+
"Qwen2-57B",
|
| 2846 |
+
"Qwen2-57B",
|
| 2847 |
+
"Qwen2-57B"
|
| 2848 |
+
],
|
| 2849 |
+
"qwen-2-57-b-instruct": [
|
| 2850 |
+
"Qwen2-57B-Inst",
|
| 2851 |
+
"Qwen2-57B-Inst",
|
| 2852 |
+
"Qwen2-57B-Inst",
|
| 2853 |
+
"Qwen2-57B-Inst",
|
| 2854 |
+
"Qwen2-57B-Inst"
|
| 2855 |
+
],
|
| 2856 |
+
"qwen-2-5-1-5-b": [
|
| 2857 |
+
"Qwen2.5-1.5B",
|
| 2858 |
+
"Qwen2.5-1.5B",
|
| 2859 |
+
"Qwen2.5-1.5B",
|
| 2860 |
+
"Qwen2.5-1.5B",
|
| 2861 |
+
"Qwen2.5-1.5B"
|
| 2862 |
+
],
|
| 2863 |
+
"qwen-2-5-1-5-b-instruct": [
|
| 2864 |
+
"Qwen2.5-1.5B-Inst",
|
| 2865 |
+
"Qwen2.5-1.5B-Inst",
|
| 2866 |
+
"Qwen2.5-1.5B-Inst",
|
| 2867 |
+
"Qwen2.5-1.5B-Inst",
|
| 2868 |
+
"Qwen2.5-1.5B-Inst"
|
| 2869 |
+
],
|
| 2870 |
+
"qwen-2-5-3-b": [
|
| 2871 |
+
"Qwen2.5-3B",
|
| 2872 |
+
"Qwen2.5-3B",
|
| 2873 |
+
"Qwen2.5-3B",
|
| 2874 |
+
"Qwen2.5-3B",
|
| 2875 |
+
"Qwen2.5-3B"
|
| 2876 |
+
],
|
| 2877 |
+
"qwen-2-5-3-b-instruct": [
|
| 2878 |
+
"Qwen2.5-3B-Inst",
|
| 2879 |
+
"Qwen2.5-3B-Inst",
|
| 2880 |
+
"Qwen2.5-3B-Inst",
|
| 2881 |
+
"Qwen2.5-3B-Inst",
|
| 2882 |
+
"Qwen2.5-3B-Inst"
|
| 2883 |
+
],
|
| 2884 |
+
"qwen-2-5-7-b": [
|
| 2885 |
+
"Qwen2.5-7B",
|
| 2886 |
+
"Qwen2.5-7B",
|
| 2887 |
+
"Qwen2.5-7B",
|
| 2888 |
+
"Qwen2.5-7B",
|
| 2889 |
+
"Qwen2.5-7B"
|
| 2890 |
+
],
|
| 2891 |
+
"qwen-2-5-72-b-instruct": [
|
| 2892 |
+
"Qwen2.5-72B-Inst",
|
| 2893 |
+
"Qwen2.5-72B-Inst",
|
| 2894 |
+
"Qwen2.5-72B-Inst",
|
| 2895 |
+
"Qwen2.5-72B-Inst",
|
| 2896 |
+
"Qwen2.5-72B-Inst"
|
| 2897 |
+
],
|
| 2898 |
+
"prolong": [
|
| 2899 |
+
"ProLong",
|
| 2900 |
+
"ProLong",
|
| 2901 |
+
"ProLong",
|
| 2902 |
+
"ProLong",
|
| 2903 |
+
"ProLong"
|
| 2904 |
+
],
|
| 2905 |
+
"gemma-3-12-b": [
|
| 2906 |
+
"Gemma-3-12B",
|
| 2907 |
+
"Gemma-3-12B",
|
| 2908 |
+
"Gemma-3-12B",
|
| 2909 |
+
"Gemma-3-12B",
|
| 2910 |
+
"Gemma-3-12B"
|
| 2911 |
+
],
|
| 2912 |
+
"gemma-3-12-b-instruct": [
|
| 2913 |
+
"Gemma-3-12B-Inst",
|
| 2914 |
+
"Gemma-3-12B-Inst",
|
| 2915 |
+
"Gemma-3-12B-Inst",
|
| 2916 |
+
"Gemma-3-12B-Inst",
|
| 2917 |
+
"Gemma-3-12B-Inst"
|
| 2918 |
+
],
|
| 2919 |
+
"gemma-3-27-b": [
|
| 2920 |
+
"Gemma-3-27B",
|
| 2921 |
+
"Gemma-3-27B",
|
| 2922 |
+
"Gemma-3-27B",
|
| 2923 |
+
"Gemma-3-27B",
|
| 2924 |
+
"Gemma-3-27B"
|
| 2925 |
+
],
|
| 2926 |
+
"gemma-3-27-b-instruct": [
|
| 2927 |
+
"Gemma-3-27B-Inst",
|
| 2928 |
+
"Gemma-3-27B-Inst",
|
| 2929 |
+
"Gemma-3-27B-Inst",
|
| 2930 |
+
"Gemma-3-27B-Inst",
|
| 2931 |
+
"Gemma-3-27B-Inst"
|
| 2932 |
+
],
|
| 2933 |
+
"jamba-v-0-1": [
|
| 2934 |
+
"Jamba-v0.1",
|
| 2935 |
+
"Jamba-v0.1",
|
| 2936 |
+
"Jamba-v0.1",
|
| 2937 |
+
"Jamba-v0.1",
|
| 2938 |
+
"Jamba-v0.1"
|
| 2939 |
+
]
|
| 2940 |
+
}
|
| 2941 |
+
}
|
data/solutions_hub.json
CHANGED
|
@@ -137,6 +137,22 @@
|
|
| 137 |
"best_for": "Predicting NIAH and reasoning pass rates from architecture alone — no inference needed.",
|
| 138 |
"not_for": "Final go/no-go decision — re-test on your domain after architectural screening passes."
|
| 139 |
},
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 140 |
{
|
| 141 |
"id": "tokenizer_glitch",
|
| 142 |
"category": "diagnostic",
|
|
|
|
| 137 |
"best_for": "Predicting NIAH and reasoning pass rates from architecture alone — no inference needed.",
|
| 138 |
"not_for": "Final go/no-go decision — re-test on your domain after architectural screening passes."
|
| 139 |
},
|
| 140 |
+
{
|
| 141 |
+
"id": "long_ctx_degradation",
|
| 142 |
+
"category": "diagnostic",
|
| 143 |
+
"pain": "Long-ctx accuracy drops well before the claimed 128K window — but raw scores are dominated by base ability and hide which model truly retains capability past short context.",
|
| 144 |
+
"tafagent_mode": "🎯 LongScore",
|
| 145 |
+
"external_tools": [
|
| 146 |
+
{"name": "100-LongBench paper (LongScore metric)", "url": "https://arxiv.org/abs/2505.19293", "type": "paper"},
|
| 147 |
+
{"name": "HELMET (Princeton, 7-task long-ctx benchmark)", "url": "https://github.com/princeton-nlp/HELMET", "type": "tool"},
|
| 148 |
+
{"name": "HELMET full results sheet (n=315)", "url": "https://docs.google.com/spreadsheets/d/1LBt6dP4UwZwU_CjoYhyAd_rjKhQLvo0Gq4cYUnpi_CA", "type": "leaderboard"},
|
| 149 |
+
{"name": "NVIDIA RULER per-length scores", "url": "https://github.com/NVIDIA/RULER", "type": "tool"},
|
| 150 |
+
{"name": "LongBench v2 (THUDM)", "url": "https://longbench2.github.io/", "type": "leaderboard"},
|
| 151 |
+
{"name": "Chroma context rot study", "url": "https://research.trychroma.com/context-rot", "type": "paper"}
|
| 152 |
+
],
|
| 153 |
+
"best_for": "Picking a long-ctx model by relative degradation rather than raw score. Uses peer-reviewed LongScore metric to disentangle base ability from long-ctx capability. Browser-only, no GPU.",
|
| 154 |
+
"not_for": "Models not yet covered by RULER or HELMET (KB has n=93). Use the external tools above for breadth."
|
| 155 |
+
},
|
| 156 |
{
|
| 157 |
"id": "tokenizer_glitch",
|
| 158 |
"category": "diagnostic",
|
index.html
CHANGED
|
@@ -231,6 +231,9 @@
|
|
| 231 |
<p><strong data-i18n="help.v087.tax.title">🌍 Multilingual Tokenizer Tax</strong></p>
|
| 232 |
<p data-i18n="help.v087.tax.body">Tokenizers tax non-English text asymmetrically. The same paragraph might be 100 tokens in English but 250+ in Chinese on a Latin-trained tokenizer (Llama, Phi). Both cost-per-request AND effective context degrade silently. This tool loads HuggingFace transformers.js in your browser (~750 KB CDN) and tokenizes pasted text against 6 preset vendor tokenizers (Qwen2.5, Phi-3.5, Llama-3.1, Gemma-2, GPT-4 cl100k, Claude approx). <em>Use case</em>: 'My multilingual support added 30% to the bill — which language costs the most?' → paste real production text, see exact per-tokenizer breakdown.</p>
|
| 233 |
|
|
|
|
|
|
|
|
|
|
| 234 |
<p><strong data-i18n="help.v081.hub.title">🧭 Solutions Hub</strong></p>
|
| 235 |
<p data-i18n="help.v081.hub.body">tafagent as integrator, not silo. 30+ pains across 7 categories (eval reliability · diagnostics · setup · training · retrieval · multimodal · observability), each mapped to (a) the tafagent mode that addresses it, if any, and (b) the best-of-breed external tools the community already trusts (RAGAS, MTEB, HELM, MCP Schema Validator, llm-stats, llguidance, GlitchMiner, etc.). Search box matches across pain, scenario, and tool name. <em>Use case</em>: 'I have problem X — does tafagent solve it, and if not, who does?'</p>
|
| 236 |
|
|
@@ -348,6 +351,7 @@
|
|
| 348 |
<li data-i18n="inv.v084.cache"><strong>🔁 Cache Diff</strong> — predicts whether a prompt edit invalidated the provider's prompt cache. Per-provider hit ratio + $ delta.</li>
|
| 349 |
<li data-i18n="inv.v085.speculative"><strong>🔬 Spec-Decode</strong> — verifies tokenizer vocab compatibility between target + draft before you ship speculative decoding (the bug that gives WORSE throughput silently).</li>
|
| 350 |
<li data-i18n="inv.v087.tax"><strong>🌍 Token Tax</strong> — real BPE encoding across 6 vendor tokenizers. Surfaces the silent cost asymmetry across languages (CJK / Arabic / mixed).</li>
|
|
|
|
| 351 |
<li data-i18n="inv.v081.hub"><strong>🧭 Solutions Hub</strong> — every documented pain mapped to a tafagent mode or curated external tool. Don't reinvent — find.</li>
|
| 352 |
</ul>
|
| 353 |
</details>
|
|
@@ -485,6 +489,7 @@
|
|
| 485 |
<button class="mode-btn" data-mode="cache" role="tab" aria-selected="false" data-i18n="modes.cache">🔁 Cache Diff</button>
|
| 486 |
<button class="mode-btn" data-mode="speculative" role="tab" aria-selected="false" data-i18n="modes.speculative">🔬 Spec-Decode</button>
|
| 487 |
<button class="mode-btn" data-mode="tax" role="tab" aria-selected="false" data-i18n="modes.tax">🌍 Token Tax</button>
|
|
|
|
| 488 |
<button class="mode-btn" data-mode="hub" role="tab" aria-selected="false" data-i18n="modes.hub">🧭 Solutions</button>
|
| 489 |
</div>
|
| 490 |
<p id="mode-desc" class="recipe-desc" data-i18n="modes.desc">
|
|
@@ -1178,6 +1183,34 @@
|
|
| 1178 |
</p>
|
| 1179 |
</section>
|
| 1180 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1181 |
<section id="hub-section" style="display:none;">
|
| 1182 |
<h2><span data-i18n="hub.title">🧭 Solutions Hub</span>
|
| 1183 |
<span class="info"><span class="tooltip" data-i18n="hub.tip">
|
|
|
|
| 231 |
<p><strong data-i18n="help.v087.tax.title">🌍 Multilingual Tokenizer Tax</strong></p>
|
| 232 |
<p data-i18n="help.v087.tax.body">Tokenizers tax non-English text asymmetrically. The same paragraph might be 100 tokens in English but 250+ in Chinese on a Latin-trained tokenizer (Llama, Phi). Both cost-per-request AND effective context degrade silently. This tool loads HuggingFace transformers.js in your browser (~750 KB CDN) and tokenizes pasted text against 6 preset vendor tokenizers (Qwen2.5, Phi-3.5, Llama-3.1, Gemma-2, GPT-4 cl100k, Claude approx). <em>Use case</em>: 'My multilingual support added 30% to the bill — which language costs the most?' → paste real production text, see exact per-tokenizer breakdown.</p>
|
| 233 |
|
| 234 |
+
<p><strong data-i18n="help.v088.longscore.title">🎯 LongScore</strong></p>
|
| 235 |
+
<p data-i18n="help.v088.longscore.body">Every long-ctx LLM claims 128K but degrades long before that. The 100-LongBench paper (ACL 2025, arXiv:2505.19293) noticed that raw long-ctx scores are dominated by base ability. They propose <strong>LongScore</strong>: <code>LC_l = (S_l − Base) / Base</code> with <code>Base = mean(S_short)</code>, then average over long lengths. This tafagent mode embeds LongScore-ready data: RULER aggregate per-context (n=33 models, 4K-128K) + HELMET aggregate at 128K (n=60 models, 7 task categories). Lookup is exact-match by HF model id. <em>Use case</em>: 'I want to use Llama-3.1-70B-Instruct for 100K-token doc summarization — how much accuracy do I actually lose?' → paste id, see -10% LongScore (moderate degradation, mostly the 128K cliff).</p>
|
| 236 |
+
|
| 237 |
<p><strong data-i18n="help.v081.hub.title">🧭 Solutions Hub</strong></p>
|
| 238 |
<p data-i18n="help.v081.hub.body">tafagent as integrator, not silo. 30+ pains across 7 categories (eval reliability · diagnostics · setup · training · retrieval · multimodal · observability), each mapped to (a) the tafagent mode that addresses it, if any, and (b) the best-of-breed external tools the community already trusts (RAGAS, MTEB, HELM, MCP Schema Validator, llm-stats, llguidance, GlitchMiner, etc.). Search box matches across pain, scenario, and tool name. <em>Use case</em>: 'I have problem X — does tafagent solve it, and if not, who does?'</p>
|
| 239 |
|
|
|
|
| 351 |
<li data-i18n="inv.v084.cache"><strong>🔁 Cache Diff</strong> — predicts whether a prompt edit invalidated the provider's prompt cache. Per-provider hit ratio + $ delta.</li>
|
| 352 |
<li data-i18n="inv.v085.speculative"><strong>🔬 Spec-Decode</strong> — verifies tokenizer vocab compatibility between target + draft before you ship speculative decoding (the bug that gives WORSE throughput silently).</li>
|
| 353 |
<li data-i18n="inv.v087.tax"><strong>🌍 Token Tax</strong> — real BPE encoding across 6 vendor tokenizers. Surfaces the silent cost asymmetry across languages (CJK / Arabic / mixed).</li>
|
| 354 |
+
<li data-i18n="inv.v088.longscore"><strong>🎯 LongScore</strong> — peer-reviewed degradation metric (100-LongBench, ACL 2025). Lookup any model in RULER + HELMET KBs (n=93). See how much your model actually drops past short context.</li>
|
| 355 |
<li data-i18n="inv.v081.hub"><strong>🧭 Solutions Hub</strong> — every documented pain mapped to a tafagent mode or curated external tool. Don't reinvent — find.</li>
|
| 356 |
</ul>
|
| 357 |
</details>
|
|
|
|
| 489 |
<button class="mode-btn" data-mode="cache" role="tab" aria-selected="false" data-i18n="modes.cache">🔁 Cache Diff</button>
|
| 490 |
<button class="mode-btn" data-mode="speculative" role="tab" aria-selected="false" data-i18n="modes.speculative">🔬 Spec-Decode</button>
|
| 491 |
<button class="mode-btn" data-mode="tax" role="tab" aria-selected="false" data-i18n="modes.tax">🌍 Token Tax</button>
|
| 492 |
+
<button class="mode-btn" data-mode="longscore" role="tab" aria-selected="false" data-i18n="modes.longscore">🎯 LongScore</button>
|
| 493 |
<button class="mode-btn" data-mode="hub" role="tab" aria-selected="false" data-i18n="modes.hub">🧭 Solutions</button>
|
| 494 |
</div>
|
| 495 |
<p id="mode-desc" class="recipe-desc" data-i18n="modes.desc">
|
|
|
|
| 1183 |
</p>
|
| 1184 |
</section>
|
| 1185 |
|
| 1186 |
+
<!-- LongScore (mode=longscore, v0.8.8 anti-bullshit pack #14) -->
|
| 1187 |
+
<section id="longscore-section" style="display:none;">
|
| 1188 |
+
<h2><span data-i18n="longscore.title">🎯 LongScore</span>
|
| 1189 |
+
<span class="info"><span class="tooltip" data-i18n="longscore.tip">
|
| 1190 |
+
<strong>Why this matters</strong>: every model claims a 128K context window, but accuracy degrades long before that. LongScore (peer-reviewed metric from 100-LongBench, ACL 2025) measures relative degradation past short context. Disentangles base ability from true long-ctx capability — so you compare degradation, not raw scores. Lookup against RULER + HELMET KBs (n=93 models).
|
| 1191 |
+
</span></span>
|
| 1192 |
+
</h2>
|
| 1193 |
+
<p class="recipe-desc" data-i18n="longscore.desc">
|
| 1194 |
+
<strong>How much does your model degrade past short context?</strong> Paste an HF model id → see LongScore (relative degradation) + per-length breakdown + HELMET 7-task scores when available. No GPU. No inference. Pure lookup against published benchmarks.
|
| 1195 |
+
</p>
|
| 1196 |
+
<div class="form-row">
|
| 1197 |
+
<label for="longscore-input" data-i18n="longscore.input_label">Model id:</label>
|
| 1198 |
+
<input type="text" id="longscore-input" data-i18n-placeholder="longscore.input.placeholder"
|
| 1199 |
+
placeholder="e.g. Qwen2.5-72B-Instruct or meta-llama/Llama-3.1-70B-Instruct" style="flex:1;" />
|
| 1200 |
+
<button type="button" id="longscore-lookup-btn" data-i18n="longscore.lookup_btn">🔎 Lookup</button>
|
| 1201 |
+
</div>
|
| 1202 |
+
<div class="form-row">
|
| 1203 |
+
<button type="button" id="longscore-example-good-btn" class="secondary" data-i18n="longscore.example_good_btn">↳ Example: Jamba-1.5-Large (no degradation)</button>
|
| 1204 |
+
<button type="button" id="longscore-example-mid-btn" class="secondary" data-i18n="longscore.example_mid_btn">↳ Example: Llama-3.1-70B (moderate)</button>
|
| 1205 |
+
<button type="button" id="longscore-example-bad-btn" class="secondary" data-i18n="longscore.example_bad_btn">↳ Example: dbrx (severe)</button>
|
| 1206 |
+
</div>
|
| 1207 |
+
<p id="longscore-status" class="recipe-desc" style="font-size:0.92em;"></p>
|
| 1208 |
+
<div id="longscore-output" style="margin-top: 1em;"></div>
|
| 1209 |
+
<p class="recipe-desc subtle" style="font-size:0.82em;margin-top:1em;" data-i18n="longscore.formula_note">
|
| 1210 |
+
💡 <strong>LongScore</strong> = mean over l ∈ {16K, 32K, 64K, 128K} of (S_l − Base) / Base, where Base = mean(S_4K, S_8K). Source: <a href="https://arxiv.org/abs/2505.19293" target="_blank">100-LongBench, ACL 2025</a>. Data: <a href="https://github.com/NVIDIA/RULER" target="_blank">NVIDIA RULER</a> (per-length, n=33) + <a href="https://github.com/princeton-nlp/HELMET" target="_blank">HELMET</a> (aggregate at 128K, n=60). 0 = no degradation; -0.30 = severe.
|
| 1211 |
+
</p>
|
| 1212 |
+
</section>
|
| 1213 |
+
|
| 1214 |
<section id="hub-section" style="display:none;">
|
| 1215 |
<h2><span data-i18n="hub.title">🧭 Solutions Hub</span>
|
| 1216 |
<span class="info"><span class="tooltip" data-i18n="hub.tip">
|
js/i18n.js
CHANGED
|
@@ -753,6 +753,44 @@ export const TRANSLATIONS = {
|
|
| 753 |
"help.v087.tax.title": "🌍 Multilingual Tokenizer Tax",
|
| 754 |
"help.v087.tax.body": "Tokenizers tax non-English text asymmetrically. The same paragraph might be 100 tokens in English but 250+ in Chinese on a Latin-trained tokenizer (Llama, Phi). Both cost-per-request AND effective context degrade silently. This tool loads HuggingFace transformers.js in your browser (~750 KB CDN) and tokenizes pasted text against 6 preset vendor tokenizers (Qwen2.5, Phi-3.5, Llama-3.1, Gemma-2, GPT-4 cl100k, Claude approx). Output: per-tokenizer token count + chars-per-token + ratio vs baseline + cost-asymmetry interpretation. Auto-detects script blocks (Latin / CJK / Arabic / Cyrillic / Devanagari / Thai / Greek / Hebrew / Korean) so users see why one tokenizer is 3× another. <em>Use case</em>: 'My multilingual support added 30% to the bill — which language costs the most?' → paste real production text, see exact per-tokenizer breakdown.",
|
| 755 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 756 |
"inv.v081.hub": "<strong>🧭 Solutions Hub</strong> — every documented pain mapped to a tafagent mode or curated external tool. Don't reinvent — find.",
|
| 757 |
"help.v081.hub.title": "🧭 Solutions Hub",
|
| 758 |
"help.v081.hub.body": "tafagent as integrator, not silo. 30+ pains across 7 categories (eval reliability · diagnostics · setup · training · retrieval · multimodal · observability), each mapped to (a) the tafagent mode that addresses it, if any, and (b) the best-of-breed external tools the community already trusts (RAGAS, MTEB, HELM, MCP Schema Validator, llm-stats, llguidance, GlitchMiner, etc.). Search box matches across pain, scenario, and tool name. <em>Use case</em>: 'I have problem X — does tafagent solve it, and if not, who does?'",
|
|
@@ -1965,6 +2003,44 @@ export const TRANSLATIONS = {
|
|
| 1965 |
"help.v087.tax.title": "🌍 Impuesto de Tokenizer Multilingüe",
|
| 1966 |
"help.v087.tax.body": "Los tokenizers gravan el texto no-inglés de forma asimétrica. El mismo párrafo puede ser 100 tokens en inglés pero 250+ en chino en un tokenizer entrenado en Latin (Llama, Phi). Tanto coste-por-request COMO contexto efectivo degradan silenciosamente. Esta tool carga HuggingFace transformers.js en tu navegador (~750 KB CDN) y tokeniza el texto pegado contra 6 tokenizers preset de vendor (Qwen2.5, Phi-3.5, Llama-3.1, Gemma-2, GPT-4 cl100k, Claude aprox). Output: token count por tokenizer + chars-per-token + ratio vs baseline + interpretación de asimetría. Auto-detecta bloques de script (Latin / CJK / árabe / cirílico / devanagari / tailandés / griego / hebreo / coreano) para que veas por qué un tokenizer es 3× otro. <em>Caso de uso</em>: 'Mi soporte multilingüe añadió 30% a la factura — ¿qué idioma cuesta más?' → pega texto real de producción, ve breakdown exacto por tokenizer.",
|
| 1967 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1968 |
"inv.v081.hub": "<strong>🧭 Solutions Hub</strong> — cada pain documentado mapeado a un mode tafagent o herramienta externa curada. No reinventes — encuentra.",
|
| 1969 |
"help.v081.hub.title": "🧭 Solutions Hub",
|
| 1970 |
"help.v081.hub.body": "tafagent como integrador, no silo. 30+ pains en 7 categorías (eval reliability · diagnósticos · setup · training · retrieval · multimodal · observability), cada uno mapeado a (a) el mode tafagent que lo resuelve, si existe, y (b) las herramientas externas best-of-breed que la comunidad ya usa (RAGAS, MTEB, HELM, MCP Schema Validator, llm-stats, llguidance, GlitchMiner, etc.). Caja de búsqueda matchea pain, scenario, y nombre de herramienta. <em>Caso de uso</em>: 'tengo problema X — ¿lo resuelve tafagent, y si no, quién?'",
|
|
@@ -3041,6 +3117,44 @@ export const TRANSLATIONS = {
|
|
| 3041 |
"help.v087.tax.title": "🌍 Taxe Tokenizer Multilingue",
|
| 3042 |
"help.v087.tax.body": "Les tokenizers taxent le texte non-anglais de façon asymétrique. Le même paragraphe peut faire 100 tokens en anglais mais 250+ en chinois sur un tokenizer entraîné en Latin (Llama, Phi). Coût-par-requête ET contexte effectif dégradent silencieusement. Cet outil charge HuggingFace transformers.js dans votre navigateur (~750 KB CDN) et tokenize le texte collé contre 6 tokenizers preset de fournisseurs (Qwen2.5, Phi-3.5, Llama-3.1, Gemma-2, GPT-4 cl100k, Claude approx). Sortie : token count par tokenizer + chars-per-token + ratio vs baseline + interprétation d'asymétrie. Auto-détecte les blocs de script (Latin / CJK / arabe / cyrillique / devanagari / thaï / grec / hébreu / coréen) pour voir pourquoi un tokenizer est 3× un autre. <em>Cas d'usage</em> : 'Mon support multilingue a ajouté 30% à la facture — quelle langue coûte le plus ?' → collez du texte de production réel, voyez le breakdown exact par tokenizer.",
|
| 3043 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 3044 |
"inv.v081.hub": "<strong>🧭 Solutions Hub</strong> — chaque pain documenté mappé à un mode tafagent ou outil externe curé. Ne réinventez pas — trouvez.",
|
| 3045 |
"help.v081.hub.title": "🧭 Solutions Hub",
|
| 3046 |
"help.v081.hub.body": "tafagent comme intégrateur, pas silo. 30+ pains à travers 7 catégories (eval reliability · diagnostics · setup · training · retrieval · multimodal · observability), chacun mappé à (a) le mode tafagent qui le résout, s'il existe, et (b) les outils externes best-of-breed que la communauté utilise déjà (RAGAS, MTEB, HELM, MCP Schema Validator, llm-stats, llguidance, GlitchMiner, etc.). La barre de recherche matche pain, scénario, et nom d'outil. <em>Cas d'usage</em> : 'j'ai le problème X — tafagent le résout-il, et sinon, qui ?'",
|
|
@@ -4117,6 +4231,44 @@ export const TRANSLATIONS = {
|
|
| 4117 |
"help.v087.tax.title": "🌍 多语言 Tokenizer 税",
|
| 4118 |
"help.v087.tax.body": "Tokenizer 对非英语文本的征税不对称。同一段落在英语中可能是 100 个 token，但在拉丁字母训练的 tokenizer（Llama、Phi）上的中文可能是 250+ 个 token。每次请求成本和有效上下文都会静默降级。这个工具在你的浏览器中加载 HuggingFace transformers.js（~750 KB CDN），并对粘贴的文本运行 6 个预设供应商 tokenizer（Qwen2.5、Phi-3.5、Llama-3.1、Gemma-2、GPT-4 cl100k、Claude 近似）的 tokenize。输出：每个 tokenizer 的 token 数 + 字符/token + 相对于 baseline 的比率 + 成本不对称解读。自动检测脚本块（拉丁/CJK/阿拉伯/西里尔/天城/泰/希腊/希伯来/韩文）让你看到为什么一个 tokenizer 是另一个的 3 倍。<em>用例</em>：『我的多语言支持给账单加了 30%——哪种语言成本最高？』→ 粘贴真实生产文本，查看每个 tokenizer 的精确分解。",
|
| 4119 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 4120 |
"inv.v081.hub": "<strong>🧭 Solutions Hub</strong> — 每个文档化的问题都映射到一个 tafagent 模式或精选外部工具。别重复发明 — 去找。",
|
| 4121 |
"help.v081.hub.title": "🧭 Solutions Hub",
|
| 4122 |
"help.v081.hub.body": "tafagent 作为集成者而非孤岛。30+ 问题跨 7 类别（评估可靠性 · 诊断 · 设置 · 训练 · 检索 · 多模态 · 可观测性），每个映射到（a）解决它的 tafagent 模式（若存在），以及（b）社区已信任的最佳外部工具（RAGAS、MTEB、HELM、MCP Schema Validator、llm-stats、llguidance、GlitchMiner 等）。搜索框匹配 pain、场景和工具名称。<em>用例</em>：'我有问题 X — tafagent 解决它吗，如果不，谁解决？'",
|
|
|
|
| 753 |
"help.v087.tax.title": "🌍 Multilingual Tokenizer Tax",
|
| 754 |
"help.v087.tax.body": "Tokenizers tax non-English text asymmetrically. The same paragraph might be 100 tokens in English but 250+ in Chinese on a Latin-trained tokenizer (Llama, Phi). Both cost-per-request AND effective context degrade silently. This tool loads HuggingFace transformers.js in your browser (~750 KB CDN) and tokenizes pasted text against 6 preset vendor tokenizers (Qwen2.5, Phi-3.5, Llama-3.1, Gemma-2, GPT-4 cl100k, Claude approx). Output: per-tokenizer token count + chars-per-token + ratio vs baseline + cost-asymmetry interpretation. Auto-detects script blocks (Latin / CJK / Arabic / Cyrillic / Devanagari / Thai / Greek / Hebrew / Korean) so users see why one tokenizer is 3× another. <em>Use case</em>: 'My multilingual support added 30% to the bill — which language costs the most?' → paste real production text, see exact per-tokenizer breakdown.",
|
| 755 |
|
| 756 |
+
// v0.8.8 — anti-bullshit pack #14: LongScore (RULER + HELMET lookup)
|
| 757 |
+
"modes.longscore": "🎯 LongScore",
|
| 758 |
+
"mode_desc.longscore": "Look up your model's relative degradation past short context. RULER + HELMET KBs (n=93 models). LongScore metric from 100-LongBench (ACL 2025).",
|
| 759 |
+
"longscore.title": "🎯 LongScore",
|
| 760 |
+
"longscore.tip": "Every model claims a 128K context window, but accuracy degrades long before that. LongScore (peer-reviewed metric from 100-LongBench, ACL 2025) measures relative degradation past short context. Disentangles base ability from true long-ctx capability — so you compare degradation, not raw scores. Lookup against RULER + HELMET KBs (n=93 models).",
|
| 761 |
+
"longscore.desc": "<strong>How much does your model degrade past short context?</strong> Paste an HF model id → see LongScore (relative degradation) + per-length breakdown + HELMET 7-task scores when available. No GPU. No inference. Pure lookup against published benchmarks.",
|
| 762 |
+
"longscore.input_label": "Model id:",
|
| 763 |
+
"longscore.input.placeholder": "e.g. Qwen2.5-72B-Instruct or meta-llama/Llama-3.1-70B-Instruct",
|
| 764 |
+
"longscore.lookup_btn": "🔎 Lookup",
|
| 765 |
+
"longscore.example_good_btn": "↳ Example: Jamba-1.5-Large (no degradation)",
|
| 766 |
+
"longscore.example_mid_btn": "↳ Example: Llama-3.1-70B (moderate)",
|
| 767 |
+
"longscore.example_bad_btn": "↳ Example: dbrx (severe)",
|
| 768 |
+
"longscore.formula_note": "💡 <strong>LongScore</strong> = mean over l ∈ {16K, 32K, 64K, 128K} of (S_l − Base) / Base, where Base = mean(S_4K, S_8K). Source: <a href=\"https://arxiv.org/abs/2505.19293\" target=\"_blank\">100-LongBench, ACL 2025</a>. Data: <a href=\"https://github.com/NVIDIA/RULER\" target=\"_blank\">NVIDIA RULER</a> (per-length, n=33) + <a href=\"https://github.com/princeton-nlp/HELMET\" target=\"_blank\">HELMET</a> (aggregate at 128K, n=60). 0 = no degradation; -0.30 = severe.",
|
| 769 |
+
"longscore.miss.title": "Model not found in KB",
|
| 770 |
+
"longscore.miss.body": "Looked up <code>{id}</code>. KB has {n} models. Try a canonical HF id (e.g. <code>Qwen2.5-72B-Instruct</code>, <code>Llama-3.1-70B-Instruct</code>, <code>Jamba-1.5-Mini</code>).",
|
| 771 |
+
"longscore.miss.suggest": "Check coverage at",
|
| 772 |
+
"longscore.no_ruler": "⚠ No per-length data — LongScore not computable. Showing HELMET aggregate at 128K instead.",
|
| 773 |
+
"longscore.score_label": "LongScore",
|
| 774 |
+
"longscore.helmet_label": "HELMET 7-task breakdown",
|
| 775 |
+
"longscore.col.ctx": "Context",
|
| 776 |
+
"longscore.col.score": "Score",
|
| 777 |
+
"longscore.col.lc": "LC",
|
| 778 |
+
"longscore.col.task": "Task",
|
| 779 |
+
"longscore.source_note": "Data source",
|
| 780 |
+
"longscore.hint.empty": "⚠ Paste a model id first.",
|
| 781 |
+
"longscore.status.lookup": "⏳ Looking up…",
|
| 782 |
+
"longscore.status.miss": "ℹ Model not in KB",
|
| 783 |
+
"longscore.status.ruler_hit": "✅ RULER per-length data found",
|
| 784 |
+
"longscore.status.helmet_only":"ℹ HELMET aggregate only (no per-length data)",
|
| 785 |
+
"longscore.verdict.no_degradation": "✅ No degradation past short context",
|
| 786 |
+
"longscore.verdict.mild": "🟢 Mild degradation (<10%)",
|
| 787 |
+
"longscore.verdict.moderate": "🟠 Moderate degradation (10-20%)",
|
| 788 |
+
"longscore.verdict.severe": "🔴 Severe degradation (20-30%)",
|
| 789 |
+
"longscore.verdict.extreme": "🚨 Extreme degradation (>30%)",
|
| 790 |
+
"inv.v088.longscore": "<strong>🎯 LongScore</strong> — peer-reviewed degradation metric (100-LongBench, ACL 2025). Lookup any model in RULER + HELMET KBs (n=93). See how much your model actually drops past short context.",
|
| 791 |
+
"help.v088.longscore.title": "🎯 LongScore",
|
| 792 |
+
"help.v088.longscore.body": "Every long-ctx LLM claims 128K but degrades long before that. The 100-LongBench paper (ACL 2025, arXiv:2505.19293) noticed that raw long-ctx scores are dominated by base ability — a smarter model with a worse long-ctx recipe still scores higher than a less-smart model with a better recipe, masking the actual long-ctx degradation. They propose <strong>LongScore</strong>: <code>LC_l = (S_l − Base) / Base</code> with <code>Base = mean(S_short)</code>, then average over long lengths. Result: a relative-degradation number per model that compares apples to apples. This tafagent mode embeds LongScore-ready data: RULER aggregate per-context (n=33 models, 4K-128K) + HELMET aggregate at 128K (n=60 models, 7 task categories). Lookup is exact-match by HF model id (lowercase, dashes, dots normalized). For models with RULER data, you get the full LongScore + per-length breakdown + verdict (no/mild/moderate/severe/extreme degradation). For HELMET-only models, you get the 7-category aggregate at 128K. <em>Use case</em>: 'I want to use Llama-3.1-70B-Instruct for 100K-token doc summarization — how much accuracy do I actually lose?' → paste id, see -10% LongScore (moderate degradation, mostly the 128K cliff). Decide whether to use it, switch to a model with engineered long-ctx, or chunk your input.",
|
| 793 |
+
|
| 794 |
"inv.v081.hub": "<strong>🧭 Solutions Hub</strong> — every documented pain mapped to a tafagent mode or curated external tool. Don't reinvent — find.",
|
| 795 |
"help.v081.hub.title": "🧭 Solutions Hub",
|
| 796 |
"help.v081.hub.body": "tafagent as integrator, not silo. 30+ pains across 7 categories (eval reliability · diagnostics · setup · training · retrieval · multimodal · observability), each mapped to (a) the tafagent mode that addresses it, if any, and (b) the best-of-breed external tools the community already trusts (RAGAS, MTEB, HELM, MCP Schema Validator, llm-stats, llguidance, GlitchMiner, etc.). Search box matches across pain, scenario, and tool name. <em>Use case</em>: 'I have problem X — does tafagent solve it, and if not, who does?'",
|
|
|
|
| 2003 |
"help.v087.tax.title": "🌍 Impuesto de Tokenizer Multilingüe",
|
| 2004 |
"help.v087.tax.body": "Los tokenizers gravan el texto no-inglés de forma asimétrica. El mismo párrafo puede ser 100 tokens en inglés pero 250+ en chino en un tokenizer entrenado en Latin (Llama, Phi). Tanto coste-por-request COMO contexto efectivo degradan silenciosamente. Esta tool carga HuggingFace transformers.js en tu navegador (~750 KB CDN) y tokeniza el texto pegado contra 6 tokenizers preset de vendor (Qwen2.5, Phi-3.5, Llama-3.1, Gemma-2, GPT-4 cl100k, Claude aprox). Output: token count por tokenizer + chars-per-token + ratio vs baseline + interpretación de asimetría. Auto-detecta bloques de script (Latin / CJK / árabe / cirílico / devanagari / tailandés / griego / hebreo / coreano) para que veas por qué un tokenizer es 3× otro. <em>Caso de uso</em>: 'Mi soporte multilingüe añadió 30% a la factura — ¿qué idioma cuesta más?' → pega texto real de producción, ve breakdown exacto por tokenizer.",
|
| 2005 |
|
| 2006 |
+
// v0.8.8 — anti-bullshit pack #14: LongScore (RULER + HELMET lookup)
|
| 2007 |
+
"modes.longscore": "🎯 LongScore",
|
| 2008 |
+
"mode_desc.longscore": "Consulta la degradación relativa de tu modelo más allá del contexto corto. KBs RULER + HELMET (n=93 modelos). Métrica LongScore de 100-LongBench (ACL 2025).",
|
| 2009 |
+
"longscore.title": "🎯 LongScore",
|
| 2010 |
+
"longscore.tip": "Cada modelo dice tener ventana de 128K, pero la accuracy degrada mucho antes. LongScore (métrica peer-reviewed de 100-LongBench, ACL 2025) mide la degradación relativa más allá del contexto corto. Separa la base ability de la capacidad real long-ctx — comparas degradación, no scores brutos. Lookup contra KBs RULER + HELMET (n=93 modelos).",
|
| 2011 |
+
"longscore.desc": "<strong>¿Cuánto degrada tu modelo más allá del contexto corto?</strong> Pega un id de modelo HF → ve LongScore (degradación relativa) + breakdown por longitud + scores HELMET 7-task cuando estén disponibles. Sin GPU. Sin inferencia. Lookup puro contra benchmarks publicados.",
|
| 2012 |
+
"longscore.input_label": "Id del modelo:",
|
| 2013 |
+
"longscore.input.placeholder": "ej. Qwen2.5-72B-Instruct o meta-llama/Llama-3.1-70B-Instruct",
|
| 2014 |
+
"longscore.lookup_btn": "🔎 Buscar",
|
| 2015 |
+
"longscore.example_good_btn": "↳ Ejemplo: Jamba-1.5-Large (sin degradación)",
|
| 2016 |
+
"longscore.example_mid_btn": "↳ Ejemplo: Llama-3.1-70B (moderado)",
|
| 2017 |
+
"longscore.example_bad_btn": "↳ Ejemplo: dbrx (severo)",
|
| 2018 |
+
"longscore.formula_note": "💡 <strong>LongScore</strong> = media sobre l ∈ {16K, 32K, 64K, 128K} de (S_l − Base) / Base, donde Base = media(S_4K, S_8K). Fuente: <a href=\"https://arxiv.org/abs/2505.19293\" target=\"_blank\">100-LongBench, ACL 2025</a>. Datos: <a href=\"https://github.com/NVIDIA/RULER\" target=\"_blank\">NVIDIA RULER</a> (per-length, n=33) + <a href=\"https://github.com/princeton-nlp/HELMET\" target=\"_blank\">HELMET</a> (agregado a 128K, n=60). 0 = sin degradación; -0.30 = severo.",
|
| 2019 |
+
"longscore.miss.title": "Modelo no encontrado en KB",
|
| 2020 |
+
"longscore.miss.body": "Buscado <code>{id}</code>. KB tiene {n} modelos. Prueba un id HF canónico (ej. <code>Qwen2.5-72B-Instruct</code>, <code>Llama-3.1-70B-Instruct</code>, <code>Jamba-1.5-Mini</code>).",
|
| 2021 |
+
"longscore.miss.suggest": "Comprueba cobertura en",
|
| 2022 |
+
"longscore.no_ruler": "⚠ Sin datos per-length — LongScore no computable. Mostrando agregado HELMET a 128K.",
|
| 2023 |
+
"longscore.score_label": "LongScore",
|
| 2024 |
+
"longscore.helmet_label": "Breakdown HELMET 7-task",
|
| 2025 |
+
"longscore.col.ctx": "Contexto",
|
| 2026 |
+
"longscore.col.score": "Score",
|
| 2027 |
+
"longscore.col.lc": "LC",
|
| 2028 |
+
"longscore.col.task": "Tarea",
|
| 2029 |
+
"longscore.source_note": "Fuente",
|
| 2030 |
+
"longscore.hint.empty": "⚠ Pega un id de modelo primero.",
|
| 2031 |
+
"longscore.status.lookup": "⏳ Buscando…",
|
| 2032 |
+
"longscore.status.miss": "ℹ Modelo no en KB",
|
| 2033 |
+
"longscore.status.ruler_hit": "✅ Datos RULER per-length encontrados",
|
| 2034 |
+
"longscore.status.helmet_only":"ℹ Solo agregado HELMET (sin datos per-length)",
|
| 2035 |
+
"longscore.verdict.no_degradation": "✅ Sin degradación más allá del contexto corto",
|
| 2036 |
+
"longscore.verdict.mild": "🟢 Degradación leve (<10%)",
|
| 2037 |
+
"longscore.verdict.moderate": "🟠 Degradación moderada (10-20%)",
|
| 2038 |
+
"longscore.verdict.severe": "🔴 Degradación severa (20-30%)",
|
| 2039 |
+
"longscore.verdict.extreme": "🚨 Degradación extrema (>30%)",
|
| 2040 |
+
"inv.v088.longscore": "<strong>🎯 LongScore</strong> — métrica de degradación peer-reviewed (100-LongBench, ACL 2025). Lookup de cualquier modelo en KBs RULER + HELMET (n=93). Ve cuánto cae tu modelo en realidad más allá del contexto corto.",
|
| 2041 |
+
"help.v088.longscore.title": "🎯 LongScore",
|
| 2042 |
+
"help.v088.longscore.body": "Cada LLM long-ctx dice 128K pero degrada mucho antes. El paper 100-LongBench (ACL 2025, arXiv:2505.19293) notó que los scores brutos long-ctx están dominados por base ability — un modelo más smart con peor receta long-ctx puntúa más que uno menos smart con mejor receta, ocultando la degradación real. Proponen <strong>LongScore</strong>: <code>LC_l = (S_l − Base) / Base</code> con <code>Base = media(S_short)</code>, luego promedio sobre longitudes largas. Resultado: número de degradación relativa por modelo que compara apples to apples. Este mode tafagent embebe datos LongScore-ready: agregado RULER per-context (n=33 modelos, 4K-128K) + agregado HELMET a 128K (n=60 modelos, 7 categorías). Lookup es match exacto por id HF (lowercase, dashes, dots normalizados). Para modelos con datos RULER, obtienes el LongScore completo + breakdown per-length + verdict (no/leve/moderado/severo/extremo). Para modelos solo-HELMET, obtienes el agregado 7-categorías a 128K. <em>Caso de uso</em>: '¿quiero usar Llama-3.1-70B-Instruct para resumen de docs 100K-token — cuánta accuracy pierdo realmente?' → pega id, ve -10% LongScore (degradación moderada, sobre todo el cliff a 128K). Decide si usarlo, cambiar a un modelo con long-ctx engineered, o chunkear tu input.",
|
| 2043 |
+
|
| 2044 |
"inv.v081.hub": "<strong>🧭 Solutions Hub</strong> — cada pain documentado mapeado a un mode tafagent o herramienta externa curada. No reinventes — encuentra.",
|
| 2045 |
"help.v081.hub.title": "🧭 Solutions Hub",
|
| 2046 |
"help.v081.hub.body": "tafagent como integrador, no silo. 30+ pains en 7 categorías (eval reliability · diagnósticos · setup · training · retrieval · multimodal · observability), cada uno mapeado a (a) el mode tafagent que lo resuelve, si existe, y (b) las herramientas externas best-of-breed que la comunidad ya usa (RAGAS, MTEB, HELM, MCP Schema Validator, llm-stats, llguidance, GlitchMiner, etc.). Caja de búsqueda matchea pain, scenario, y nombre de herramienta. <em>Caso de uso</em>: 'tengo problema X — ¿lo resuelve tafagent, y si no, quién?'",
|
|
|
|
| 3117 |
"help.v087.tax.title": "🌍 Taxe Tokenizer Multilingue",
|
| 3118 |
"help.v087.tax.body": "Les tokenizers taxent le texte non-anglais de façon asymétrique. Le même paragraphe peut faire 100 tokens en anglais mais 250+ en chinois sur un tokenizer entraîné en Latin (Llama, Phi). Coût-par-requête ET contexte effectif dégradent silencieusement. Cet outil charge HuggingFace transformers.js dans votre navigateur (~750 KB CDN) et tokenize le texte collé contre 6 tokenizers preset de fournisseurs (Qwen2.5, Phi-3.5, Llama-3.1, Gemma-2, GPT-4 cl100k, Claude approx). Sortie : token count par tokenizer + chars-per-token + ratio vs baseline + interprétation d'asymétrie. Auto-détecte les blocs de script (Latin / CJK / arabe / cyrillique / devanagari / thaï / grec / hébreu / coréen) pour voir pourquoi un tokenizer est 3× un autre. <em>Cas d'usage</em> : 'Mon support multilingue a ajouté 30% à la facture — quelle langue coûte le plus ?' → collez du texte de production réel, voyez le breakdown exact par tokenizer.",
|
| 3119 |
|
| 3120 |
+
// v0.8.8 — anti-bullshit pack #14 : LongScore (RULER + HELMET lookup)
|
| 3121 |
+
"modes.longscore": "���� LongScore",
|
| 3122 |
+
"mode_desc.longscore": "Recherchez la dégradation relative de votre modèle au-delà du contexte court. KBs RULER + HELMET (n=93 modèles). Métrique LongScore de 100-LongBench (ACL 2025).",
|
| 3123 |
+
"longscore.title": "🎯 LongScore",
|
| 3124 |
+
"longscore.tip": "Chaque modèle prétend une fenêtre 128K, mais la précision dégrade bien avant. LongScore (métrique peer-reviewed de 100-LongBench, ACL 2025) mesure la dégradation relative au-delà du contexte court. Sépare la base ability de la vraie capacité long-ctx — vous comparez la dégradation, pas les scores bruts. Lookup contre KBs RULER + HELMET (n=93 modèles).",
|
| 3125 |
+
"longscore.desc": "<strong>Combien votre modèle dégrade-t-il au-delà du contexte court ?</strong> Collez un id modèle HF → voyez LongScore (dégradation relative) + breakdown par longueur + scores HELMET 7-task quand disponibles. Pas de GPU. Pas d'inférence. Lookup pur contre des benchmarks publiés.",
|
| 3126 |
+
"longscore.input_label": "Id du modèle :",
|
| 3127 |
+
"longscore.input.placeholder": "ex. Qwen2.5-72B-Instruct ou meta-llama/Llama-3.1-70B-Instruct",
|
| 3128 |
+
"longscore.lookup_btn": "🔎 Rechercher",
|
| 3129 |
+
"longscore.example_good_btn": "↳ Exemple : Jamba-1.5-Large (sans dégradation)",
|
| 3130 |
+
"longscore.example_mid_btn": "↳ Exemple : Llama-3.1-70B (modéré)",
|
| 3131 |
+
"longscore.example_bad_btn": "↳ Exemple : dbrx (sévère)",
|
| 3132 |
+
"longscore.formula_note": "💡 <strong>LongScore</strong> = moyenne sur l ∈ {16K, 32K, 64K, 128K} de (S_l − Base) / Base, où Base = moyenne(S_4K, S_8K). Source : <a href=\"https://arxiv.org/abs/2505.19293\" target=\"_blank\">100-LongBench, ACL 2025</a>. Données : <a href=\"https://github.com/NVIDIA/RULER\" target=\"_blank\">NVIDIA RULER</a> (per-length, n=33) + <a href=\"https://github.com/princeton-nlp/HELMET\" target=\"_blank\">HELMET</a> (agrégat à 128K, n=60). 0 = pas de dégradation ; -0.30 = sévère.",
|
| 3133 |
+
"longscore.miss.title": "Modèle non trouvé en KB",
|
| 3134 |
+
"longscore.miss.body": "Recherché <code>{id}</code>. KB contient {n} modèles. Essayez un id HF canonique (ex. <code>Qwen2.5-72B-Instruct</code>, <code>Llama-3.1-70B-Instruct</code>, <code>Jamba-1.5-Mini</code>).",
|
| 3135 |
+
"longscore.miss.suggest": "Vérifiez la couverture sur",
|
| 3136 |
+
"longscore.no_ruler": "⚠ Pas de données per-length — LongScore non calculable. Affichage agrégat HELMET à 128K.",
|
| 3137 |
+
"longscore.score_label": "LongScore",
|
| 3138 |
+
"longscore.helmet_label": "Breakdown HELMET 7-task",
|
| 3139 |
+
"longscore.col.ctx": "Contexte",
|
| 3140 |
+
"longscore.col.score": "Score",
|
| 3141 |
+
"longscore.col.lc": "LC",
|
| 3142 |
+
"longscore.col.task": "Tâche",
|
| 3143 |
+
"longscore.source_note": "Source",
|
| 3144 |
+
"longscore.hint.empty": "⚠ Collez un id modèle d'abord.",
|
| 3145 |
+
"longscore.status.lookup": "⏳ Recherche…",
|
| 3146 |
+
"longscore.status.miss": "ℹ Modèle pas en KB",
|
| 3147 |
+
"longscore.status.ruler_hit": "✅ Données RULER per-length trouvées",
|
| 3148 |
+
"longscore.status.helmet_only":"ℹ Agrégat HELMET seulement (pas de données per-length)",
|
| 3149 |
+
"longscore.verdict.no_degradation": "✅ Pas de dégradation au-delà du contexte court",
|
| 3150 |
+
"longscore.verdict.mild": "🟢 Dégradation légère (<10%)",
|
| 3151 |
+
"longscore.verdict.moderate": "🟠 Dégradation modérée (10-20%)",
|
| 3152 |
+
"longscore.verdict.severe": "🔴 Dégradation sévère (20-30%)",
|
| 3153 |
+
"longscore.verdict.extreme": "🚨 Dégradation extrême (>30%)",
|
| 3154 |
+
"inv.v088.longscore": "<strong>🎯 LongScore</strong> — métrique de dégradation peer-reviewed (100-LongBench, ACL 2025). Lookup de tout modèle dans KBs RULER + HELMET (n=93). Voyez combien votre modèle chute réellement au-delà du contexte court.",
|
| 3155 |
+
"help.v088.longscore.title": "🎯 LongScore",
|
| 3156 |
+
"help.v088.longscore.body": "Chaque LLM long-ctx prétend 128K mais dégrade bien avant. Le paper 100-LongBench (ACL 2025, arXiv:2505.19293) a remarqué que les scores long-ctx bruts sont dominés par la base ability — un modèle plus smart avec une moins bonne recette long-ctx score plus qu'un moins smart avec une meilleure recette, masquant la vraie dégradation. Ils proposent <strong>LongScore</strong> : <code>LC_l = (S_l − Base) / Base</code> avec <code>Base = moyenne(S_short)</code>, puis moyenne sur les longueurs longues. Résultat : un nombre de dégradation relative par modèle qui compare apples to apples. Ce mode tafagent embarque les données LongScore-ready : agrégat RULER per-context (n=33 modèles, 4K-128K) + agrégat HELMET à 128K (n=60 modèles, 7 catégories). Lookup est match exact par id HF (lowercase, dashes, dots normalisés). Pour les modèles avec données RULER, vous obtenez le LongScore complet + breakdown per-length + verdict (pas/légère/modérée/sévère/extrême). Pour les modèles HELMET-only, vous obtenez l'agrégat 7-catégories à 128K. <em>Cas d'usage</em> : 'je veux utiliser Llama-3.1-70B-Instruct pour résumé de docs 100K-token — combien de précision je perds vraiment ?' → collez l'id, voyez -10% LongScore (dégradation modérée, surtout le cliff à 128K). Décidez de l'utiliser, passer à un modèle avec long-ctx engineered, ou chunker votre input.",
|
| 3157 |
+
|
| 3158 |
"inv.v081.hub": "<strong>🧭 Solutions Hub</strong> — chaque pain documenté mappé à un mode tafagent ou outil externe curé. Ne réinventez pas — trouvez.",
|
| 3159 |
"help.v081.hub.title": "🧭 Solutions Hub",
|
| 3160 |
"help.v081.hub.body": "tafagent comme intégrateur, pas silo. 30+ pains à travers 7 catégories (eval reliability · diagnostics · setup · training · retrieval · multimodal · observability), chacun mappé à (a) le mode tafagent qui le résout, s'il existe, et (b) les outils externes best-of-breed que la communauté utilise déjà (RAGAS, MTEB, HELM, MCP Schema Validator, llm-stats, llguidance, GlitchMiner, etc.). La barre de recherche matche pain, scénario, et nom d'outil. <em>Cas d'usage</em> : 'j'ai le problème X — tafagent le résout-il, et sinon, qui ?'",
|
|
|
|
| 4231 |
"help.v087.tax.title": "🌍 多语言 Tokenizer 税",
|
| 4232 |
"help.v087.tax.body": "Tokenizer 对非英语文本的征税不对称。同一段落在英语中可能是 100 个 token，但在拉丁字母训练的 tokenizer（Llama、Phi）上的中文可能是 250+ 个 token。每次请求成本和有效上下文都会静默降级。这个工具在你的浏览器中加载 HuggingFace transformers.js（~750 KB CDN），并对粘贴的文本运行 6 个预设供应商 tokenizer（Qwen2.5、Phi-3.5、Llama-3.1、Gemma-2、GPT-4 cl100k、Claude 近似）的 tokenize。输出：每个 tokenizer 的 token 数 + 字符/token + 相对于 baseline 的比率 + 成本不对称解读。自动检测脚本块（拉丁/CJK/阿拉伯/西里尔/天城/泰/希腊/希伯来/韩文）让你看到为什么一个 tokenizer 是另一个的 3 倍。<em>用例</em>：『我的多语言支持给账单加了 30%——哪种语言成本最高？』→ 粘贴真实生产文本，查看每个 tokenizer 的精确分解。",
|
| 4233 |
|
| 4234 |
+
// v0.8.8 — anti-bullshit pack #14：LongScore (RULER + HELMET 查询)
|
| 4235 |
+
"modes.longscore": "🎯 LongScore",
|
| 4236 |
+
"mode_desc.longscore": "查询你的模型在短上下文之外的相对降级。RULER + HELMET KB（n=93 模型）。LongScore 指标来自 100-LongBench (ACL 2025)。",
|
| 4237 |
+
"longscore.title": "🎯 LongScore",
|
| 4238 |
+
"longscore.tip": "每个模型都声称 128K 上下文窗口，但准确率早就开始降级。LongScore（来自 100-LongBench, ACL 2025 的 peer-reviewed 指标）测量相对于短上下文的降级。将基础能力与真正的长上下文能力解耦——你比较的是降级，而不是原始分数。在 RULER + HELMET KB 中查询（n=93 模型）。",
|
| 4239 |
+
"longscore.desc": "<strong>你的模型在短上下文之外降级多少？</strong> 粘贴 HF 模型 id → 查看 LongScore（相对降级）+ 每长度分解 + HELMET 7-task 分数（如有）。无 GPU。无推理。纯查询已发布的 benchmark。",
|
| 4240 |
+
"longscore.input_label": "模型 id：",
|
| 4241 |
+
"longscore.input.placeholder": "例如 Qwen2.5-72B-Instruct 或 meta-llama/Llama-3.1-70B-Instruct",
|
| 4242 |
+
"longscore.lookup_btn": "🔎 查询",
|
| 4243 |
+
"longscore.example_good_btn": "↳ 示例：Jamba-1.5-Large（无降级）",
|
| 4244 |
+
"longscore.example_mid_btn": "↳ 示例：Llama-3.1-70B（中等）",
|
| 4245 |
+
"longscore.example_bad_btn": "↳ 示例：dbrx（严重）",
|
| 4246 |
+
"longscore.formula_note": "💡 <strong>LongScore</strong> = 在 l ∈ {16K, 32K, 64K, 128K} 上的 (S_l − Base) / Base 平均值，其中 Base = mean(S_4K, S_8K)。来源：<a href=\"https://arxiv.org/abs/2505.19293\" target=\"_blank\">100-LongBench, ACL 2025</a>。数据：<a href=\"https://github.com/NVIDIA/RULER\" target=\"_blank\">NVIDIA RULER</a>（每长度，n=33）+ <a href=\"https://github.com/princeton-nlp/HELMET\" target=\"_blank\">HELMET</a>（128K 聚合，n=60）。0 = 无降级；-0.30 = 严重。",
|
| 4247 |
+
"longscore.miss.title": "KB 中未找到模型",
|
| 4248 |
+
"longscore.miss.body": "查询了 <code>{id}</code>。KB 包含 {n} 个模型。请尝试规范 HF id（例如 <code>Qwen2.5-72B-Instruct</code>、<code>Llama-3.1-70B-Instruct</code>、<code>Jamba-1.5-Mini</code>）。",
|
| 4249 |
+
"longscore.miss.suggest": "在以下位置检查覆盖范围",
|
| 4250 |
+
"longscore.no_ruler": "⚠ 无每长度数据 — LongScore 无法计算。改为显示 128K 处的 HELMET 聚合。",
|
| 4251 |
+
"longscore.score_label": "LongScore",
|
| 4252 |
+
"longscore.helmet_label": "HELMET 7-task 分解",
|
| 4253 |
+
"longscore.col.ctx": "上下文",
|
| 4254 |
+
"longscore.col.score": "分数",
|
| 4255 |
+
"longscore.col.lc": "LC",
|
| 4256 |
+
"longscore.col.task": "任务",
|
| 4257 |
+
"longscore.source_note": "数据源",
|
| 4258 |
+
"longscore.hint.empty": "⚠ 请先粘贴模型 id。",
|
| 4259 |
+
"longscore.status.lookup": "⏳ 查询中…",
|
| 4260 |
+
"longscore.status.miss": "ℹ 模型不在 KB 中",
|
| 4261 |
+
"longscore.status.ruler_hit": "✅ 找到 RULER 每长度数据",
|
| 4262 |
+
"longscore.status.helmet_only":"ℹ 仅 HELMET 聚合（无每长度数据）",
|
| 4263 |
+
"longscore.verdict.no_degradation": "✅ 短上下文之外无降级",
|
| 4264 |
+
"longscore.verdict.mild": "🟢 轻度降级（<10%）",
|
| 4265 |
+
"longscore.verdict.moderate": "🟠 中度降级（10-20%）",
|
| 4266 |
+
"longscore.verdict.severe": "🔴 严重降级（20-30%）",
|
| 4267 |
+
"longscore.verdict.extreme": "🚨 极端降级（>30%）",
|
| 4268 |
+
"inv.v088.longscore": "<strong>🎯 LongScore</strong> — peer-reviewed 降级指标（100-LongBench, ACL 2025）。在 RULER + HELMET KB（n=93）中查询任意模型。看你的模型在短上下文之外实际下降多少。",
|
| 4269 |
+
"help.v088.longscore.title": "🎯 LongScore",
|
| 4270 |
+
"help.v088.longscore.body": "每个长上下文 LLM 都声称 128K，但早就开始降级。100-LongBench 论文 (ACL 2025, arXiv:2505.19293) 注意到原始长上下文分数被基础能力主导——一个更聪明但长上下文配方更差的模型，仍然得分高于一个不那么聪明但配方更好的模型，掩盖了真正的长上下文降级。他们提出 <strong>LongScore</strong>：<code>LC_l = (S_l − Base) / Base</code>，其中 <code>Base = mean(S_short)</code>，然后对长长度取平均。结果：每个模型一个相对降级数字，可以同等比较。这个 tafagent 模式嵌入了 LongScore-ready 数据：RULER 每上下文聚合（n=33 模型，4K-128K）+ HELMET 128K 聚合（n=60 模型，7 类别）。查询是按 HF 模型 id 精确匹配（小写、连字符、点号已规范化）。对于有 RULER 数据的模型，你得到完整的 LongScore + 每长度分解 + 判定（无/轻/中/严重/极端降级）。对于仅 HELMET 模型，你得到 128K 处的 7-类别聚合。<em>用例</em>：『我想用 Llama-3.1-70B-Instruct 做 100K-token 文档摘要——实际上我损失多少准确率？』→ 粘贴 id，看到 -10% LongScore（中度降级，主要是 128K 处的 cliff）。决定是否使用、改用 long-ctx engineered 的模型，或者分块输入。",
|
| 4271 |
+
|
| 4272 |
"inv.v081.hub": "<strong>🧭 Solutions Hub</strong> — 每个文档化的问题都映射到一个 tafagent 模式或精选外部工具。别重复发明 — 去找。",
|
| 4273 |
"help.v081.hub.title": "🧭 Solutions Hub",
|
| 4274 |
"help.v081.hub.body": "tafagent 作为集成者而非孤岛。30+ 问题跨 7 类别（评估可靠性 · 诊断 · 设置 · 训练 · 检索 · 多模态 · 可观测性），每个映射到（a）解决它的 tafagent 模式（若存在），以及（b）社区已信任的最佳外部工具（RAGAS、MTEB、HELM、MCP Schema Validator、llm-stats、llguidance、GlitchMiner 等）。搜索框匹配 pain、场景和工具名称。<em>用例</em>：'我有问题 X — tafagent 解决它吗，如果不，谁解决？'",
|
js/longscore.js
ADDED
|
@@ -0,0 +1,105 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
// longscore.js — pure logic for the 🎯 LongScore mode.
|
| 2 |
+
//
|
| 3 |
+
// Looks up an HF-style model id in data/longscore_kb.json and returns:
|
| 4 |
+
// - exact match: ruler_per_ctx (if available) + ruler_long_score (computed) + helmet aggregate
|
| 5 |
+
// - HELMET-only: aggregate scores at 128K, no LongScore (no per-length data)
|
| 6 |
+
// - miss: fallback for unknown models
|
| 7 |
+
//
|
| 8 |
+
// No UI strings — emits codes + params; main.js translates via i18n.
|
| 9 |
+
//
|
| 10 |
+
// LongScore formula (100-LongBench, ACL 2025, arXiv:2505.19293, §3.2):
|
| 11 |
+
// Base = mean(S_4K, S_8K)
|
| 12 |
+
// LC_l = (S_l - Base) / Base
|
| 13 |
+
// LongScore = mean(LC_l for l in {16K, 32K, 64K, 128K})
|
| 14 |
+
//
|
| 15 |
+
// More negative = worse long-ctx retention.
|
| 16 |
+
|
| 17 |
+
let KB = null;
|
| 18 |
+
|
| 19 |
+
export async function loadKB() {
|
| 20 |
+
if (KB) return KB;
|
| 21 |
+
const res = await fetch("data/longscore_kb.json");
|
| 22 |
+
if (!res.ok) throw new Error("longscore_kb fetch failed: " + res.status);
|
| 23 |
+
KB = await res.json();
|
| 24 |
+
return KB;
|
| 25 |
+
}
|
| 26 |
+
|
| 27 |
+
export function normalize(name) {
|
| 28 |
+
if (!name) return "";
|
| 29 |
+
let s = String(name).toLowerCase().trim();
|
| 30 |
+
s = s.replace(/^(meta-llama\/|01-ai\/|ai21labs\/|nvidia\/|princeton-nlp\/|unsloth\/)/, "");
|
| 31 |
+
s = s.replace(/_/g, "-").replace(/\./g, "-");
|
| 32 |
+
s = s.replace(/([a-z])(\d)/g, "$1-$2");
|
| 33 |
+
s = s.replace(/(\d)([a-z])/g, "$1-$2");
|
| 34 |
+
s = s.replace(/-+/g, "-");
|
| 35 |
+
// -inst → -instruct (both at end and in middle, before next -segment)
|
| 36 |
+
s = s.replace(/-inst(?=-|$)/g, "-instruct");
|
| 37 |
+
return s;
|
| 38 |
+
}
|
| 39 |
+
|
| 40 |
+
/** Classify LongScore avg into verdict code. */
|
| 41 |
+
export function classify(longscore_avg, thresholds) {
|
| 42 |
+
if (longscore_avg === null || longscore_avg === undefined) return "no_data";
|
| 43 |
+
if (longscore_avg >= thresholds.no_degradation) return "no_degradation";
|
| 44 |
+
if (longscore_avg >= thresholds.mild) return "mild";
|
| 45 |
+
if (longscore_avg >= thresholds.moderate) return "moderate";
|
| 46 |
+
if (longscore_avg >= thresholds.severe) return "severe";
|
| 47 |
+
return "extreme";
|
| 48 |
+
}
|
| 49 |
+
|
| 50 |
+
/** Look up a model and return a structured result. */
|
| 51 |
+
export async function lookup(rawId) {
|
| 52 |
+
const kb = await loadKB();
|
| 53 |
+
const id = normalize(rawId);
|
| 54 |
+
const entry = kb.models[id];
|
| 55 |
+
if (!entry) {
|
| 56 |
+
return {
|
| 57 |
+
code: "miss",
|
| 58 |
+
normalized_id: id,
|
| 59 |
+
n_kb_total: kb.stats.n_total,
|
| 60 |
+
};
|
| 61 |
+
}
|
| 62 |
+
|
| 63 |
+
const longscore = entry.ruler_long_score;
|
| 64 |
+
const verdict = longscore
|
| 65 |
+
? classify(longscore.avg_lc, kb.thresholds)
|
| 66 |
+
: null;
|
| 67 |
+
|
| 68 |
+
return {
|
| 69 |
+
code: longscore ? "ruler_hit" : (entry.helmet ? "helmet_only" : "partial"),
|
| 70 |
+
display_name: entry.display_name,
|
| 71 |
+
normalized_id: id,
|
| 72 |
+
ruler_per_ctx: entry.ruler_per_ctx,
|
| 73 |
+
ruler_long_score: longscore,
|
| 74 |
+
helmet: entry.helmet,
|
| 75 |
+
recipe_class: entry.recipe_class,
|
| 76 |
+
params_b: entry.params_b,
|
| 77 |
+
native_context_k: entry.native_context_k,
|
| 78 |
+
source: entry.source,
|
| 79 |
+
verdict,
|
| 80 |
+
thresholds: kb.thresholds,
|
| 81 |
+
};
|
| 82 |
+
}
|
| 83 |
+
|
| 84 |
+
/** Get sorted list of all model ids — for autocomplete. */
|
| 85 |
+
export async function listAllIds() {
|
| 86 |
+
const kb = await loadKB();
|
| 87 |
+
return Object.keys(kb.models).sort();
|
| 88 |
+
}
|
| 89 |
+
|
| 90 |
+
/** Top-N best/worst by LongScore (for sanity inspection). Optional helper. */
|
| 91 |
+
export async function rank(direction) {
|
| 92 |
+
const kb = await loadKB();
|
| 93 |
+
const items = Object.entries(kb.models)
|
| 94 |
+
.filter(([, m]) => m.ruler_long_score)
|
| 95 |
+
.map(([id, m]) => ({
|
| 96 |
+
id,
|
| 97 |
+
display_name: m.display_name,
|
| 98 |
+
recipe_class: m.recipe_class,
|
| 99 |
+
avg_lc: m.ruler_long_score.avg_lc,
|
| 100 |
+
}));
|
| 101 |
+
items.sort((a, b) =>
|
| 102 |
+
direction === "best" ? b.avg_lc - a.avg_lc : a.avg_lc - b.avg_lc
|
| 103 |
+
);
|
| 104 |
+
return items;
|
| 105 |
+
}
|
js/main.js
CHANGED
|
@@ -35,6 +35,9 @@ import {
|
|
| 35 |
tokenizeAll, detectLanguageBlocks,
|
| 36 |
PRESET_TOKENIZERS as TAX_PRESETS, SAMPLE_TEXTS as TAX_SAMPLES,
|
| 37 |
} from "./tokenizer_tax.js";
|
|
|
|
|
|
|
|
|
|
| 38 |
|
| 39 |
// Attach HF Hub search-as-you-type to all 5 model id inputs (Profile, Recipe,
|
| 40 |
// Unmask, Template, Quant). Hits public huggingface.co/api/models. Idempotent.
|
|
@@ -229,6 +232,7 @@ document.addEventListener("click", (e) => {
|
|
| 229 |
cache: "cache-section",
|
| 230 |
speculative: "speculative-section",
|
| 231 |
tax: "tax-section",
|
|
|
|
| 232 |
hub: "hub-section",
|
| 233 |
}[targetMode];
|
| 234 |
if (sectionId) {
|
|
@@ -254,7 +258,7 @@ document.querySelectorAll(".mode-btn").forEach(btn => {
|
|
| 254 |
"diagnose-section", "phase-section", "unmask-section",
|
| 255 |
"template-section", "arena-section", "contam-section",
|
| 256 |
"quant-section", "drift-section", "niah-section",
|
| 257 |
-
"saturation-section", "cot-section", "peft-section", "cache-section", "speculative-section", "tax-section", "hub-section"].forEach(id => {
|
| 258 |
const el = $(id);
|
| 259 |
if (el) el.style.display = "none";
|
| 260 |
});
|
|
@@ -271,6 +275,7 @@ document.querySelectorAll(".mode-btn").forEach(btn => {
|
|
| 271 |
cache: "cache-section",
|
| 272 |
speculative: "speculative-section",
|
| 273 |
tax: "tax-section",
|
|
|
|
| 274 |
hub: "hub-section",
|
| 275 |
};
|
| 276 |
const sectionId = sectionMap[mode];
|
|
@@ -283,6 +288,7 @@ document.querySelectorAll(".mode-btn").forEach(btn => {
|
|
| 283 |
if (mode === "cache") initCacheDiff();
|
| 284 |
if (mode === "speculative") initSpeculative();
|
| 285 |
if (mode === "tax") initTax();
|
|
|
|
| 286 |
if (mode === "hub") initHub();
|
| 287 |
});
|
| 288 |
});
|
|
@@ -3469,10 +3475,10 @@ async function initHub() {
|
|
| 3469 |
|
| 3470 |
function renderEntry(e) {
|
| 3471 |
const modeBadge = e.tafagent_mode
|
| 3472 |
-
? `<span class="badge" style="background:#3fb950;">${e.tafagent_mode}</span>`
|
| 3473 |
: (e.tafagent_planned_mode
|
| 3474 |
-
? `<span class="badge" style="background:#d29922;">${t("hub.planned") || "planned:"} ${e.tafagent_planned_mode}</span>`
|
| 3475 |
-
: `<span class="badge" style="background:#6e7781;">${t("hub.no_mode") || "external"}</span>`);
|
| 3476 |
const tools = (e.external_tools || [])
|
| 3477 |
.map(tl => {
|
| 3478 |
const icon = HUB_TYPE_BADGE[tl.type] || "🔗";
|
|
@@ -4410,6 +4416,167 @@ $("tax-sample-code-btn")?.addEventListener("click", () => {
|
|
| 4410 |
runTaxTokenize();
|
| 4411 |
});
|
| 4412 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 4413 |
// ════════════════════════════════════════════════════════════════════
|
| 4414 |
// Bootstrap
|
| 4415 |
// ════════════════════════════════════════════════════════════════════
|
|
|
|
| 35 |
tokenizeAll, detectLanguageBlocks,
|
| 36 |
PRESET_TOKENIZERS as TAX_PRESETS, SAMPLE_TEXTS as TAX_SAMPLES,
|
| 37 |
} from "./tokenizer_tax.js";
|
| 38 |
+
import {
|
| 39 |
+
loadKB as loadLongscoreKB, lookup as longscoreLookup, rank as longscoreRank,
|
| 40 |
+
} from "./longscore.js";
|
| 41 |
|
| 42 |
// Attach HF Hub search-as-you-type to all 5 model id inputs (Profile, Recipe,
|
| 43 |
// Unmask, Template, Quant). Hits public huggingface.co/api/models. Idempotent.
|
|
|
|
| 232 |
cache: "cache-section",
|
| 233 |
speculative: "speculative-section",
|
| 234 |
tax: "tax-section",
|
| 235 |
+
longscore: "longscore-section",
|
| 236 |
hub: "hub-section",
|
| 237 |
}[targetMode];
|
| 238 |
if (sectionId) {
|
|
|
|
| 258 |
"diagnose-section", "phase-section", "unmask-section",
|
| 259 |
"template-section", "arena-section", "contam-section",
|
| 260 |
"quant-section", "drift-section", "niah-section",
|
| 261 |
+
"saturation-section", "cot-section", "peft-section", "cache-section", "speculative-section", "tax-section", "longscore-section", "hub-section"].forEach(id => {
|
| 262 |
const el = $(id);
|
| 263 |
if (el) el.style.display = "none";
|
| 264 |
});
|
|
|
|
| 275 |
cache: "cache-section",
|
| 276 |
speculative: "speculative-section",
|
| 277 |
tax: "tax-section",
|
| 278 |
+
longscore: "longscore-section",
|
| 279 |
hub: "hub-section",
|
| 280 |
};
|
| 281 |
const sectionId = sectionMap[mode];
|
|
|
|
| 288 |
if (mode === "cache") initCacheDiff();
|
| 289 |
if (mode === "speculative") initSpeculative();
|
| 290 |
if (mode === "tax") initTax();
|
| 291 |
+
if (mode === "longscore") initLongscore();
|
| 292 |
if (mode === "hub") initHub();
|
| 293 |
});
|
| 294 |
});
|
|
|
|
| 3475 |
|
| 3476 |
function renderEntry(e) {
|
| 3477 |
const modeBadge = e.tafagent_mode
|
| 3478 |
+
? `<span class="badge" style="background:#3fb950;color:#fff;border-color:#3fb950;">${e.tafagent_mode}</span>`
|
| 3479 |
: (e.tafagent_planned_mode
|
| 3480 |
+
? `<span class="badge" style="background:#d29922;color:#1a1a1a;border-color:#d29922;">${t("hub.planned") || "planned:"} ${e.tafagent_planned_mode}</span>`
|
| 3481 |
+
: `<span class="badge" style="background:#6e7781;color:#fff;border-color:#6e7781;">${t("hub.no_mode") || "external"}</span>`);
|
| 3482 |
const tools = (e.external_tools || [])
|
| 3483 |
.map(tl => {
|
| 3484 |
const icon = HUB_TYPE_BADGE[tl.type] || "🔗";
|
|
|
|
| 4416 |
runTaxTokenize();
|
| 4417 |
});
|
| 4418 |
|
| 4419 |
+
// ════════════════════════════════════════════════════════════════════
|
| 4420 |
+
// LongScore mode (v0.8.8 anti-bullshit pack #14)
|
| 4421 |
+
// ════════════════════════════════════════════════════════════════════
|
| 4422 |
+
let __longscoreInited = false;
|
| 4423 |
+
|
| 4424 |
+
function initLongscore() {
|
| 4425 |
+
if (__longscoreInited) return;
|
| 4426 |
+
__longscoreInited = true;
|
| 4427 |
+
// Eager-load KB so the first lookup is instant (KB is ~70KB, no real cost)
|
| 4428 |
+
loadLongscoreKB().catch(e => {
|
| 4429 |
+
console.warn("longscore_kb preload failed", e);
|
| 4430 |
+
});
|
| 4431 |
+
}
|
| 4432 |
+
|
| 4433 |
+
function fmtPct(x, sign) {
|
| 4434 |
+
if (x == null) return "—";
|
| 4435 |
+
const v = (x * 100);
|
| 4436 |
+
return `${sign && v >= 0 ? "+" : ""}${v.toFixed(1)}%`;
|
| 4437 |
+
}
|
| 4438 |
+
|
| 4439 |
+
function lcColor(avg) {
|
| 4440 |
+
if (avg == null) return "#8b949e";
|
| 4441 |
+
if (avg >= -0.02) return "#3fb950"; // green: no degradation
|
| 4442 |
+
if (avg >= -0.10) return "#a5d36a"; // light green
|
| 4443 |
+
if (avg >= -0.20) return "#f0883e"; // orange
|
| 4444 |
+
if (avg >= -0.30) return "#f85149"; // red
|
| 4445 |
+
return "#a01b1b"; // dark red: extreme
|
| 4446 |
+
}
|
| 4447 |
+
|
| 4448 |
+
function renderLongscoreResult(res) {
|
| 4449 |
+
if (res.code === "miss") {
|
| 4450 |
+
return `<div class="arena-result">
|
| 4451 |
+
<p style="color:#f0883e;"><strong>${t("longscore.miss.title") || "Model not found in KB"}</strong></p>
|
| 4452 |
+
<p>${tFmt("longscore.miss.body", { id: res.normalized_id, n: res.n_kb_total }) || `Looked up <code>${res.normalized_id}</code>. KB has ${res.n_kb_total} models. Try a canonical HF id (e.g. <code>Qwen2.5-72B-Instruct</code>, <code>Llama-3.1-70B-Instruct</code>, <code>Jamba-1.5-Mini</code>).`}</p>
|
| 4453 |
+
<p class="subtle" style="font-size:0.85em;">${t("longscore.miss.suggest") || "Check coverage at"} <a href="https://github.com/NVIDIA/RULER" target="_blank">RULER</a> · <a href="https://github.com/princeton-nlp/HELMET" target="_blank">HELMET</a>.</p>
|
| 4454 |
+
</div>`;
|
| 4455 |
+
}
|
| 4456 |
+
|
| 4457 |
+
const verdictMap = {
|
| 4458 |
+
no_degradation: { color: "#3fb950", label: t("longscore.verdict.no_degradation") || "✅ No degradation past short context" },
|
| 4459 |
+
mild: { color: "#a5d36a", label: t("longscore.verdict.mild") || "🟢 Mild degradation (<10%)" },
|
| 4460 |
+
moderate: { color: "#f0883e", label: t("longscore.verdict.moderate") || "🟠 Moderate degradation (10-20%)" },
|
| 4461 |
+
severe: { color: "#f85149", label: t("longscore.verdict.severe") || "🔴 Severe degradation (20-30%)" },
|
| 4462 |
+
extreme: { color: "#a01b1b", label: t("longscore.verdict.extreme") || "🚨 Extreme degradation (>30%)" },
|
| 4463 |
+
};
|
| 4464 |
+
|
| 4465 |
+
let html = `<div class="arena-result">`;
|
| 4466 |
+
html += `<p><strong>${escapeHtml(res.display_name)}</strong>`;
|
| 4467 |
+
if (res.params_b) html += ` <span class="subtle">· ${res.params_b}B params</span>`;
|
| 4468 |
+
if (res.recipe_class) html += ` <span class="subtle">· ${escapeHtml(res.recipe_class)}</span>`;
|
| 4469 |
+
if (res.native_context_k) html += ` <span class="subtle">· native ctx ${res.native_context_k}K</span>`;
|
| 4470 |
+
html += `</p>`;
|
| 4471 |
+
|
| 4472 |
+
// RULER per-length + LongScore
|
| 4473 |
+
if (res.ruler_long_score) {
|
| 4474 |
+
const ls = res.ruler_long_score;
|
| 4475 |
+
const v = verdictMap[res.verdict] || { color: "#8b949e", label: res.verdict };
|
| 4476 |
+
html += `<p style="margin-top:0.8em;font-size:1.1em;">
|
| 4477 |
+
<strong>${t("longscore.score_label") || "LongScore"}:</strong>
|
| 4478 |
+
<span style="color:${lcColor(ls.avg_lc)};font-family:monospace;font-size:1.2em;font-weight:bold;">${fmtPct(ls.avg_lc, true)}</span>
|
| 4479 |
+
<span class="subtle">· Base = ${ls.base.toFixed(1)}% (mean of 4K, 8K)</span>
|
| 4480 |
+
</p>`;
|
| 4481 |
+
html += `<p style="color:${v.color};font-weight:bold;">${v.label}</p>`;
|
| 4482 |
+
|
| 4483 |
+
// Per-length bars
|
| 4484 |
+
html += `<table class="lean-table" style="margin-top:0.8em;width:100%;">
|
| 4485 |
+
<thead><tr>
|
| 4486 |
+
<th style="text-align:left;">${t("longscore.col.ctx") || "Context"}</th>
|
| 4487 |
+
<th style="text-align:right;">${t("longscore.col.score") || "Score"}</th>
|
| 4488 |
+
<th style="text-align:right;">${t("longscore.col.lc") || "LC"}</th>
|
| 4489 |
+
</tr></thead><tbody>`;
|
| 4490 |
+
const ctxKeys = ["4k", "8k", "16k", "32k", "64k", "128k"];
|
| 4491 |
+
for (const k of ctxKeys) {
|
| 4492 |
+
const score = res.ruler_per_ctx?.[k];
|
| 4493 |
+
if (score == null) continue;
|
| 4494 |
+
const isShort = k === "4k" || k === "8k";
|
| 4495 |
+
const lc = ls.per_length_lc?.[k];
|
| 4496 |
+
html += `<tr ${isShort ? 'style="opacity:0.7;"' : ""}>
|
| 4497 |
+
<td><strong>${k.toUpperCase()}</strong>${isShort ? ` <span class="subtle" style="font-size:0.8em;">(base)</span>` : ""}</td>
|
| 4498 |
+
<td style="text-align:right;font-family:monospace;">${score.toFixed(1)}%</td>
|
| 4499 |
+
<td style="text-align:right;font-family:monospace;color:${lcColor(lc)};">${lc != null ? fmtPct(lc, true) : "—"}</td>
|
| 4500 |
+
</tr>`;
|
| 4501 |
+
}
|
| 4502 |
+
html += `</tbody></table>`;
|
| 4503 |
+
} else {
|
| 4504 |
+
// Helmet-only or partial
|
| 4505 |
+
html += `<p style="margin-top:0.8em;color:#f0883e;">${t("longscore.no_ruler") || "⚠ No per-length data — LongScore not computable. Showing HELMET aggregate at 128K instead."}</p>`;
|
| 4506 |
+
}
|
| 4507 |
+
|
| 4508 |
+
// HELMET breakdown if available
|
| 4509 |
+
if (res.helmet) {
|
| 4510 |
+
html += `<details style="margin-top:1em;" open>
|
| 4511 |
+
<summary><strong>${t("longscore.helmet_label") || "HELMET 7-task breakdown"} (at 128K)</strong></summary>
|
| 4512 |
+
<table class="lean-table" style="margin-top:0.5em;width:100%;">
|
| 4513 |
+
<thead><tr>
|
| 4514 |
+
<th style="text-align:left;">${t("longscore.col.task") || "Task"}</th>
|
| 4515 |
+
<th style="text-align:right;">${t("longscore.col.score") || "Score"}</th>
|
| 4516 |
+
</tr></thead><tbody>`;
|
| 4517 |
+
if (res.helmet.overall != null) {
|
| 4518 |
+
html += `<tr style="background:#1f2933;"><td><strong>Overall</strong></td><td style="text-align:right;font-family:monospace;"><strong>${res.helmet.overall.toFixed(1)}</strong></td></tr>`;
|
| 4519 |
+
}
|
| 4520 |
+
if (res.helmet.categories) {
|
| 4521 |
+
for (const [task, score] of Object.entries(res.helmet.categories)) {
|
| 4522 |
+
html += `<tr><td>${escapeHtml(task)}</td><td style="text-align:right;font-family:monospace;">${score != null ? score.toFixed(1) : "—"}</td></tr>`;
|
| 4523 |
+
}
|
| 4524 |
+
}
|
| 4525 |
+
html += `</tbody></table></details>`;
|
| 4526 |
+
}
|
| 4527 |
+
|
| 4528 |
+
html += `<p class="recipe-desc subtle" style="font-size:0.82em;margin-top:1em;">
|
| 4529 |
+
${t("longscore.source_note") || "Data source"}: ${escapeHtml(res.source)} ·
|
| 4530 |
+
<a href="https://arxiv.org/abs/2505.19293" target="_blank">LongScore metric</a>
|
| 4531 |
+
</p>`;
|
| 4532 |
+
html += `</div>`;
|
| 4533 |
+
return html;
|
| 4534 |
+
}
|
| 4535 |
+
|
| 4536 |
+
async function runLongscoreLookup() {
|
| 4537 |
+
const id = $("longscore-input")?.value?.trim();
|
| 4538 |
+
if (!id) {
|
| 4539 |
+
$("longscore-status").textContent = t("longscore.hint.empty") || "⚠ Paste a model id first.";
|
| 4540 |
+
return;
|
| 4541 |
+
}
|
| 4542 |
+
$("longscore-status").textContent = t("longscore.status.lookup") || "⏳ Looking up…";
|
| 4543 |
+
$("longscore-output").innerHTML = "";
|
| 4544 |
+
try {
|
| 4545 |
+
const res = await longscoreLookup(id);
|
| 4546 |
+
$("longscore-output").innerHTML = renderLongscoreResult(res);
|
| 4547 |
+
if (res.code === "miss") {
|
| 4548 |
+
$("longscore-status").textContent = t("longscore.status.miss") || "ℹ Model not in KB";
|
| 4549 |
+
} else if (res.code === "ruler_hit") {
|
| 4550 |
+
$("longscore-status").textContent = t("longscore.status.ruler_hit") || "✅ RULER per-length data found";
|
| 4551 |
+
} else {
|
| 4552 |
+
$("longscore-status").textContent = t("longscore.status.helmet_only") || "ℹ HELMET aggregate only (no per-length data)";
|
| 4553 |
+
}
|
| 4554 |
+
} catch (e) {
|
| 4555 |
+
$("longscore-status").textContent = `❌ ${e.message || e}`;
|
| 4556 |
+
console.error(e);
|
| 4557 |
+
}
|
| 4558 |
+
}
|
| 4559 |
+
|
| 4560 |
+
$("longscore-lookup-btn")?.addEventListener("click", runLongscoreLookup);
|
| 4561 |
+
$("longscore-input")?.addEventListener("keydown", e => {
|
| 4562 |
+
if (e.key === "Enter") {
|
| 4563 |
+
e.preventDefault();
|
| 4564 |
+
runLongscoreLookup();
|
| 4565 |
+
}
|
| 4566 |
+
});
|
| 4567 |
+
$("longscore-example-good-btn")?.addEventListener("click", () => {
|
| 4568 |
+
$("longscore-input").value = "Jamba-1.5-Large";
|
| 4569 |
+
runLongscoreLookup();
|
| 4570 |
+
});
|
| 4571 |
+
$("longscore-example-mid-btn")?.addEventListener("click", () => {
|
| 4572 |
+
$("longscore-input").value = "Llama-3.1-70B-Instruct";
|
| 4573 |
+
runLongscoreLookup();
|
| 4574 |
+
});
|
| 4575 |
+
$("longscore-example-bad-btn")?.addEventListener("click", () => {
|
| 4576 |
+
$("longscore-input").value = "dbrx";
|
| 4577 |
+
runLongscoreLookup();
|
| 4578 |
+
});
|
| 4579 |
+
|
| 4580 |
// ════════════════════════════════════════════════════════════════════
|
| 4581 |
// Bootstrap
|
| 4582 |
// ════════════════════════════════════════════════════════════════════
|
scripts/test_longscore.mjs
ADDED
|
@@ -0,0 +1,72 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
// Smoke test for js/longscore.js — verifies normalize, lookup, classify codes.
|
| 2 |
+
// Run: node scripts/test_longscore.mjs
|
| 3 |
+
import { readFileSync } from "fs";
|
| 4 |
+
|
| 5 |
+
// Mock fetch for Node ESM
|
| 6 |
+
globalThis.fetch = async (url) => {
|
| 7 |
+
const path = url.startsWith("data/") ? `./${url}` : url;
|
| 8 |
+
const txt = readFileSync(path, "utf-8");
|
| 9 |
+
return {
|
| 10 |
+
ok: true,
|
| 11 |
+
json: async () => JSON.parse(txt),
|
| 12 |
+
};
|
| 13 |
+
};
|
| 14 |
+
|
| 15 |
+
const { normalize, lookup, classify, rank } = await import("../js/longscore.js");
|
| 16 |
+
|
| 17 |
+
let pass = 0, fail = 0;
|
| 18 |
+
function check(name, cond, detail) {
|
| 19 |
+
if (cond) { pass++; console.log(` ✓ ${name}`); }
|
| 20 |
+
else { fail++; console.log(` ✗ ${name}${detail ? ": " + detail : ""}`); }
|
| 21 |
+
}
|
| 22 |
+
|
| 23 |
+
console.log("--- normalize ---");
|
| 24 |
+
check("trims + lowercases", normalize(" Qwen2.5 ") === "qwen-2-5");
|
| 25 |
+
check("strips meta-llama/", normalize("meta-llama/Llama-3.1-70B-Instruct") === "llama-3-1-70-b-instruct");
|
| 26 |
+
check("strips 01-ai/", normalize("01-ai/Yi-34B-200K") === "yi-34-b-200-k");
|
| 27 |
+
check("inst → instruct", normalize("Mistral-7B-Inst-v0.2") === "mistral-7-b-instruct-v-0-2");
|
| 28 |
+
check("dot → dash", normalize("Phi-3.5-mini-instruct") === "phi-3-5-mini-instruct");
|
| 29 |
+
check("empty", normalize("") === "");
|
| 30 |
+
|
| 31 |
+
console.log("\n--- classify ---");
|
| 32 |
+
const t = { no_degradation: -0.02, mild: -0.10, moderate: -0.20, severe: -0.30 };
|
| 33 |
+
check("no_data", classify(null, t) === "no_data");
|
| 34 |
+
check("no_degradation", classify(0.0, t) === "no_degradation");
|
| 35 |
+
check("mild", classify(-0.05, t) === "mild");
|
| 36 |
+
check("moderate", classify(-0.15, t) === "moderate");
|
| 37 |
+
check("severe", classify(-0.25, t) === "severe");
|
| 38 |
+
check("extreme", classify(-0.50, t) === "extreme");
|
| 39 |
+
|
| 40 |
+
console.log("\n--- lookup (RULER hit) ---");
|
| 41 |
+
const r1 = await lookup("Llama-3.1-70B-Instruct");
|
| 42 |
+
check("ruler_hit code", r1.code === "ruler_hit");
|
| 43 |
+
check("longscore present", typeof r1.ruler_long_score?.avg_lc === "number");
|
| 44 |
+
check("verdict assigned", r1.verdict !== null);
|
| 45 |
+
check("base ~96", r1.ruler_long_score?.base > 95 && r1.ruler_long_score?.base < 97,
|
| 46 |
+
`got base=${r1.ruler_long_score?.base}`);
|
| 47 |
+
check("Llama-3.1-70B avg_lc ~-0.10", Math.abs(r1.ruler_long_score?.avg_lc - (-0.1024)) < 0.001,
|
| 48 |
+
`got ${r1.ruler_long_score?.avg_lc}`);
|
| 49 |
+
|
| 50 |
+
console.log("\n--- lookup (Jamba — best LongScore) ---");
|
| 51 |
+
const r2 = await lookup("Jamba-1.5-Large");
|
| 52 |
+
check("ruler_hit", r2.code === "ruler_hit");
|
| 53 |
+
check("Jamba near-zero degradation", r2.ruler_long_score?.avg_lc > -0.02);
|
| 54 |
+
|
| 55 |
+
console.log("\n--- lookup (dbrx — severe) ---");
|
| 56 |
+
const r3 = await lookup("dbrx");
|
| 57 |
+
check("ruler_hit", r3.code === "ruler_hit");
|
| 58 |
+
check("dbrx severe verdict", r3.verdict === "severe" || r3.verdict === "extreme",
|
| 59 |
+
`got verdict=${r3.verdict} for avg_lc=${r3.ruler_long_score?.avg_lc}`);
|
| 60 |
+
|
| 61 |
+
console.log("\n--- lookup (miss) ---");
|
| 62 |
+
const r4 = await lookup("nonexistent-model-123");
|
| 63 |
+
check("miss code", r4.code === "miss");
|
| 64 |
+
check("normalized id present", r4.normalized_id === "nonexistent-model-123");
|
| 65 |
+
|
| 66 |
+
console.log("\n--- rank ---");
|
| 67 |
+
const ranking = await rank("worst");
|
| 68 |
+
check("ranking returned", Array.isArray(ranking) && ranking.length > 0);
|
| 69 |
+
check("worst is most negative", ranking[0].avg_lc < ranking[ranking.length - 1].avg_lc);
|
| 70 |
+
|
| 71 |
+
console.log(`\n${pass} passed, ${fail} failed`);
|
| 72 |
+
process.exit(fail > 0 ? 1 : 0);
|
scripts/test_longscore_e2e.mjs
ADDED
|
@@ -0,0 +1,34 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
// E2E lookup smoke for the 3 example buttons (Jamba/Llama/dbrx) + a HELMET-only model.
|
| 2 |
+
import { readFileSync } from "fs";
|
| 3 |
+
globalThis.fetch = async (url) => {
|
| 4 |
+
const path = url.startsWith("data/") ? `./${url}` : url;
|
| 5 |
+
return { ok: true, json: async () => JSON.parse(readFileSync(path, "utf-8")) };
|
| 6 |
+
};
|
| 7 |
+
|
| 8 |
+
const { lookup } = await import("../js/longscore.js");
|
| 9 |
+
|
| 10 |
+
const cases = [
|
| 11 |
+
{ input: "Jamba-1.5-Large", expect: { code: "ruler_hit", verdict: "no_degradation" } },
|
| 12 |
+
{ input: "Llama-3.1-70B-Instruct", expect: { code: "ruler_hit", verdict: "moderate" } },
|
| 13 |
+
{ input: "dbrx", expect: { code: "ruler_hit", verdict: "extreme" } },
|
| 14 |
+
{ input: "GPT-4", expect: { code: "helmet_only" } }, // HELMET-only
|
| 15 |
+
{ input: "totally-fake-model-xyz", expect: { code: "miss" } },
|
| 16 |
+
];
|
| 17 |
+
|
| 18 |
+
let pass = 0, fail = 0;
|
| 19 |
+
for (const c of cases) {
|
| 20 |
+
const r = await lookup(c.input);
|
| 21 |
+
const ok = r.code === c.expect.code &&
|
| 22 |
+
(!c.expect.verdict || r.verdict === c.expect.verdict);
|
| 23 |
+
if (ok) {
|
| 24 |
+
pass++;
|
| 25 |
+
const score = r.ruler_long_score ? `LongScore=${(r.ruler_long_score.avg_lc*100).toFixed(1)}%` :
|
| 26 |
+
r.helmet ? `HELMET overall=${r.helmet.overall}` : "";
|
| 27 |
+
console.log(` ✓ ${c.input.padEnd(30)} → ${r.code.padEnd(12)} ${r.verdict || "n/a".padEnd(15)} ${score}`);
|
| 28 |
+
} else {
|
| 29 |
+
fail++;
|
| 30 |
+
console.log(` ✗ ${c.input.padEnd(30)} → got code=${r.code} verdict=${r.verdict}, expected=${JSON.stringify(c.expect)}`);
|
| 31 |
+
}
|
| 32 |
+
}
|
| 33 |
+
console.log(`\n${pass}/${pass+fail} cases pass`);
|
| 34 |
+
process.exit(fail > 0 ? 1 : 0);
|