Spaces:
Sleeping
Sleeping
Pablo Claude Opus 4.7 (1M context) commited on
Commit ·
cf99605
1
Parent(s): da68829
docs: wire 5 dashboard/benchmark screenshots into README
Browse filesLive Demo section: dashboard_live_demo.png (With vs Without comparison).
V6 Live Snapshot block: dashboard_v6_snapshot.png + dashboard_aiter_config.png.
Benchmark section: benchmark_v6_terminal.png under the 8/8 targets table.
Key Stats section: dashboard_key_stats.png next to the stats table.
assets/screenshots/.gitkeep documents the 5 expected filenames.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
- README.md +25 -13
- assets/screenshots/.gitkeep +9 -0
README.md
CHANGED
|
@@ -82,21 +82,11 @@ Every optimization traces back to a peer-reviewed paper published at **NeurIPS,
|
|
| 82 |
|
| 83 |
## 🎬 Live Demo
|
| 84 |
|
| 85 |
-
Real metrics from `demo/app.py` running against the full ContextForge stack — five agents, real Qwen3 tokenizer, real LSH+FAISS dedup, INV-15 enforced live.
|
| 86 |
|
| 87 |
<p align="center">
|
| 88 |
-
<img src="assets/screenshots/
|
| 89 |
-
<em>Live Demo tab —
|
| 90 |
-
</p>
|
| 91 |
-
|
| 92 |
-
<p align="center">
|
| 93 |
-
<img src="assets/screenshots/dashboard_results.png" alt="Live Demo with ContextForge — 79.85% savings, INV-15 firing" width="900"><br>
|
| 94 |
-
<em>With ContextForge: <b>263 → 53 tokens (79.85 % savings)</b>, JCR Safety Gate fires INV-15 on the Critic.</em>
|
| 95 |
-
</p>
|
| 96 |
-
|
| 97 |
-
<p align="center">
|
| 98 |
-
<img src="assets/screenshots/dashboard_v6_snapshot.png" alt="Architecture tab — V6 Live Snapshot" width="900"><br>
|
| 99 |
-
<em>Architecture tab — TokenDance + JCR Safety Gate + AITER ROCm config snapshots.</em>
|
| 100 |
</p>
|
| 101 |
|
| 102 |
```
|
|
@@ -118,6 +108,18 @@ strategy: register+lsh+faiss
|
|
| 118 |
reason: INV-15: judge role='critic' risk=1.00 > threshold=0.70 → dense prefill mandated
|
| 119 |
```
|
| 120 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 121 |
---
|
| 122 |
|
| 123 |
## 🏗️ Architecture
|
|
@@ -204,6 +206,11 @@ flowchart TB
|
|
| 204 |
| JCR INV-15 violations | **0** | 0 | ✅ |
|
| 205 |
| JCR Critic dense rate (high-risk sweep) | **1.000** | ≥ 0.5 | ✅ |
|
| 206 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 207 |
---
|
| 208 |
|
| 209 |
## 📈 Key Stats
|
|
@@ -220,6 +227,11 @@ flowchart TB
|
|
| 220 |
| Peer-reviewed papers implemented | **10** |
|
| 221 |
| System invariants enforced | **15** |
|
| 222 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 223 |
---
|
| 224 |
|
| 225 |
## 🚀 Quick Start
|
|
|
|
| 82 |
|
| 83 |
## 🎬 Live Demo
|
| 84 |
|
| 85 |
+
Real metrics from `demo/app.py` running against the full ContextForge stack — five agents, real Qwen3 tokenizer, real LSH+FAISS dedup, INV-15 enforced live. Side-by-side comparison: **263 → 53 tokens, 79.85 % savings** with ContextForge; passthrough on the right.
|
| 86 |
|
| 87 |
<p align="center">
|
| 88 |
+
<img src="assets/screenshots/dashboard_live_demo.png" alt="Live Demo — With vs Without ContextForge, 79.85% savings, INV-15 firing on the Critic" width="960"><br>
|
| 89 |
+
<em>Live Demo tab — left: query input. Right: <b>With ContextForge</b> (79.85 % savings, INV-15 fires on the Critic) vs. <b>Without ContextForge</b> (passthrough, 0 % savings).</em>
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 90 |
</p>
|
| 91 |
|
| 92 |
```
|
|
|
|
| 108 |
reason: INV-15: judge role='critic' risk=1.00 > threshold=0.70 → dense prefill mandated
|
| 109 |
```
|
| 110 |
|
| 111 |
+
### V6 Live Snapshot — TokenDance + JCR Safety Gate
|
| 112 |
+
|
| 113 |
+
<p align="center">
|
| 114 |
+
<img src="assets/screenshots/dashboard_v6_snapshot.png" alt="Architecture tab — TokenDance Master-Mirror + JCR Safety Gate live snapshots" width="640"><br>
|
| 115 |
+
<em>Architecture tab — <b>TokenDance Master-Mirror Storage</b> (5-agent demo, 4.71× compression) and <b>JCR Safety Gate</b> firing INV-15 (risk = 1.000, dense_prefill = True).</em>
|
| 116 |
+
</p>
|
| 117 |
+
|
| 118 |
+
<p align="center">
|
| 119 |
+
<img src="assets/screenshots/dashboard_aiter_config.png" alt="AITER ROCm Config — MI300X" width="520"><br>
|
| 120 |
+
<em>AITER ROCm Config (MI300X) — <code>rocm_available: True</code>, 7 documented env vars, AMD-published speedups: 3× fused MoE, 2× block-scale GEMM, 2-4× FP8 memory.</em>
|
| 121 |
+
</p>
|
| 122 |
+
|
| 123 |
---
|
| 124 |
|
| 125 |
## 🏗️ Architecture
|
|
|
|
| 206 |
| JCR INV-15 violations | **0** | 0 | ✅ |
|
| 207 |
| JCR Critic dense rate (high-risk sweep) | **1.000** | ≥ 0.5 | ✅ |
|
| 208 |
|
| 209 |
+
<p align="center">
|
| 210 |
+
<img src="assets/screenshots/benchmark_v6_terminal.png" alt="V6.0 benchmark terminal output — S-14 token_dance 10.81x, S-15 jcr_gate 0 violations" width="780"><br>
|
| 211 |
+
<em>Live terminal output of <code>python demo/benchmark_v5.py</code> — S-14 TokenDance <b>10.81×</b> compression with reconstruction error <b>1.19e-07</b>, S-15 JCR Safety Gate <b>0 INV-15 violations</b>.</em>
|
| 212 |
+
</p>
|
| 213 |
+
|
| 214 |
---
|
| 215 |
|
| 216 |
## 📈 Key Stats
|
|
|
|
| 227 |
| Peer-reviewed papers implemented | **10** |
|
| 228 |
| System invariants enforced | **15** |
|
| 229 |
|
| 230 |
+
<p align="center">
|
| 231 |
+
<img src="assets/screenshots/dashboard_key_stats.png" alt="Architecture tab — Key Statistics panel" width="520"><br>
|
| 232 |
+
<em>Key Statistics panel rendered live in the dashboard's Architecture tab.</em>
|
| 233 |
+
</p>
|
| 234 |
+
|
| 235 |
---
|
| 236 |
|
| 237 |
## 🚀 Quick Start
|
assets/screenshots/.gitkeep
ADDED
|
@@ -0,0 +1,9 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# Drop the 5 dashboard / benchmark screenshots in this directory:
|
| 2 |
+
#
|
| 3 |
+
# dashboard_live_demo.png — Live Demo tab (With vs Without ContextForge, 79.85% savings)
|
| 4 |
+
# dashboard_v6_snapshot.png — Architecture tab V6 Live Snapshot (TokenDance + JCR Safety Gate)
|
| 5 |
+
# dashboard_aiter_config.png — AITER ROCm Config table + documented speedups
|
| 6 |
+
# dashboard_key_stats.png — Key Statistics panel
|
| 7 |
+
# benchmark_v6_terminal.png — V6.0 METRICS terminal output (S-14 + S-15)
|
| 8 |
+
#
|
| 9 |
+
# README.md references them at these exact paths.
|