Spaces:
Running
Running
v2 report part 1: Jackrong methodology + Darwin heretic
Browse files- RESEARCH_REPORT_v2.md +48 -0
RESEARCH_REPORT_v2.md
ADDED
|
@@ -0,0 +1,48 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# NEXUS OS Research Report v2 2026-05-21
|
| 2 |
+
|
| 3 |
+
## 1 GGUF Download Fix
|
| 4 |
+
WRONG: hf download mradermacher/Darwin-2B-Opus-GGUF:Q6_K --local-dir .models
|
| 5 |
+
RIGHT: huggingface-cli download mradermacher/Darwin-2B-Opus-GGUF --include *.Q6_K.gguf --local-dir .models
|
| 6 |
+
|
| 7 |
+
## 2 Darwin Space Fix
|
| 8 |
+
PR CREATED: https://huggingface.co/spaces/build-small-hackathon/Darwin-35B-A3B-Opus/discussions/1
|
| 9 |
+
|
| 10 |
+
## 3 Darwin Model Family
|
| 11 |
+
| Model | Params | Base | Trait |
|
| 12 |
+
|------------|--------|------------|--------------------------|
|
| 13 |
+
| Darwin-2B | 1.9B | Qwen3.5-2B | Claude-Opus distilled |
|
| 14 |
+
| Darwin-4B | 4B | Gemma-4 | MRI-guided DARE-TIES |
|
| 15 |
+
| Darwin-27B | 27B | Qwen3.6 | Evolutionary merge |
|
| 16 |
+
| Darwin-31B | 31B | Gemma-4 | Delphi reasoning |
|
| 17 |
+
| Darwin-35B | 35B | Qwen3.6 MoE| 3B active / 35B total |
|
| 18 |
+
| Darwin-36B | 36B | Qwen3.6 | Apex, bench leader |
|
| 19 |
+
| lastbrain | 1.9B | Qwen3.5-2B | SFT+LoRA merged |
|
| 20 |
+
| Darwin-2B-heretic | 1.9B | Qwen3.5 | UNcensored, abliterated |
|
| 21 |
+
|
| 22 |
+
## 4 Jackrong Methodology Summary
|
| 23 |
+
| Technique | What it is | Evidence |
|
| 24 |
+
|-----------|------------|----------|
|
| 25 |
+
| Trace Inversion | Reconstruct hidden CoT from API-only model | 9K + 5K datasets |
|
| 26 |
+
| Negentropy | Quality filtering by information density | Negentropy-based grading |
|
| 27 |
+
| Hermes Agent Traces | Multi-turn tool trajectories from Kimi + GLM-5.1 | 15K total |
|
| 28 |
+
| Scale | 934K downloads on 27B with only 3.7K training samples! | Proof: quality > quantity |
|
| 29 |
+
|
| 30 |
+
## 5 ZeroGPU Canonical Pattern
|
| 31 |
+
From build-small-hackathon/GRM-2.6-Opus:
|
| 32 |
+
- @spaces.GPU(duration=fn, size='large')
|
| 33 |
+
- BitsAndBytesConfig: bnb_4bit_quant_type='nf4'
|
| 34 |
+
- TextIteratorStreamer for streaming
|
| 35 |
+
- Thinking blocks with
|
| 36 |
+
|
| 37 |
+
## 6 Dataset Matrix v2
|
| 38 |
+
| Dataset | Rows | Format | Priority | Use |
|
| 39 |
+
|---------|------|--------|----------|-----|
|
| 40 |
+
| Jackrong/Claude-opus-4.7-TraceInversion-5000x | 5K | messages + thinking | HIGHEST | Agent SFT + reasoning |
|
| 41 |
+
| Jackrong/Claude-opus-4.6-TraceInversion-9000x | 9K | messages + thinking | HIGHEST | Agent SFT + reasoning |
|
| 42 |
+
| lambda/hermes-agent-reasoning-traces | 15K | conversations + tools | HIGHEST | Agent tool use |
|
| 43 |
+
| Jackrong/Qwen3.5-reasoning-700x | 700 | conversation | HIGH | Deep reasoning |
|
| 44 |
+
| OpenThoughts-114k | 114K | conversations | HIGH | Breadth SFT |
|
| 45 |
+
| nohurry/Opus-4.6-Reasoning-3000x | 3K | problem + thinking + solution | HIGH | Math reasoning |
|
| 46 |
+
| OpenThoughts-Agent-v1-SFT | 104K | agent traces | MEDIUM | Agent specialist |
|
| 47 |
+
| FINAL-Bench/Metacognitive | 100 | TICOS eval tasks | MEDIUM | Evaluation |
|
| 48 |
+
| FINAL-Bench/World-Model | 100 | embodied AI tasks | LOW | Evaluation |
|