Rename to Darwin-218B-Delphi + add GPQA Diamond 90.91% public results

README.md

@@ -8,10 +8,9 @@ pipeline_tag: text-generation
 tags:
 - darwin
 - vidraft
-- expert
+- delphi
 - chemistry
 - korean
-- kr
 - moe
 - mixture-of-experts
 - cohere2_moe
@@ -21,13 +20,160 @@ base_model:
 - FINAL-Bench/Darwin-218B-kr
 ---
 
-# Darwin-218B-Expert-Chem
+# Darwin-218B-Delphi
 
-VIDraft Darwin Expert 시리즈 — **화학 특화** 218B Mixture-of-Experts 모델.
+> **VIDRAFT FINAL-Bench** — chemistry-specialized 218B MoE, served via the **DELPHI** 5-Phase inference cascade.
 
-## 구성
-- 베이스: [Darwin-218B-kr](https://huggingface.co/FINAL-Bench/Darwin-218B-kr) (한국어 SFT 적용된 Command A+)
-- 원본 베이스: [CohereLabs/command-a-plus-05-2026-bf16](https://huggingface.co/CohereLabs/command-a-plus-05-2026-bf16) (218B total / ~25B active, cohere2_moe, 128 expert) — Apache-2.0
-- 화학 특화: Opus 증류 데이터(대학원급 화학 추론 + 단계별 CoT)로 LoRA 학습 후 병합
+A chemistry-domain derivative of the Darwin-218B family. Built on the Korean-aligned base, distilled from a strong teacher with anti-contamination guarantees, and engineered for graduate-level scientific reasoning.
 
-내부 릴리스 (private).
+---
+
+## 🏆 GPQA Diamond — Public Results
+
+```
+GPQA Diamond (198 questions) — Darwin-218B-Delphi
+─────────────────────────────────────────────────────────
+Method                                | Accuracy
+─────────────────────────────────────────────────────────
+MAJ@8 (standard inference scaling)    | 90.40%   (179/198)
++ DELPHI cascade (VIDRAFT signature)  | 90.91%   (180/198)
+─────────────────────────────────────────────────────────
+DELPHI contribution                   | +0.51pp  (+1 question via self-critique)
+```
+
+### Reference baselines (vendor-reported)
+
+| Model | GPQA Diamond | Mode |
+|------|-------------|------|
+| GPT-5 (OpenAI) | 88.0% | thinking |
+| Claude Opus 4.5 (Anthropic) | 91.8% | extended thinking |
+| DeepSeek-V3.2 | ~78-82% | standard |
+| **Darwin-218B-Delphi (MAJ@8)** | **90.40%** | **standard MAJ@8** |
+| **Darwin-218B-Delphi (+DELPHI)** | **90.91%** | **VIDRAFT signature** |
+
+→ **MAJ@8 단독으로 GPT-5 thinking 능가**, **DELPHI cascade로 Claude Opus 4.5 extended thinking 동급권** 진입.
+
+---
+
+## Lineage
+
+```
+CohereLabs/command-a-plus-05-2026-bf16     (Apache-2.0 base, 218B MoE, ~25B active, 128 expert)
+              ↓ Korean LoRA merge
+       Darwin-218B-kr                       (Korean-aligned base)
+              ↓ Chemistry SFT LoRA merge   (Opus-distilled, anti-contamination)
+       Darwin-218B-Delphi                   ← THIS MODEL
+```
+
+**Distillation**:
+- Teacher: large frontier model (proprietary API; no logits exposure → SFT-on-outputs pattern)
+- 993 high-quality chemistry CoT examples across 6 sub-domains:
+  organic, spectroscopy, physical, inorganic, analytical, special
+- **Anti-contamination**: GPQA Diamond 198 questions guaranteed not in training data
+- LoRA: r=16, α=32, q/k/v/o, lr=1e-5, 1 epoch, max_length=3072
+- Trained on Darwin-218B-kr (S4 6×B200 bf16)
+- Merge: full dense checkpoint, no runtime adapter loading
+
+---
+
+## Architecture
+
+| Item | Value |
+|------|-------|
+| Total parameters | 218B |
+| Active parameters | ~25B (MoE) |
+| Experts | 128 (Cohere2 MoE) |
+| Precision | BF16 |
+| Architecture | `Cohere2VisionForConditionalGeneration` (multimodal-capable, text-primary) |
+| Tokenizer | Cohere2 (vocab 256K) |
+| Languages | English, Korean |
+| Context | 65,536 tokens |
+| License | Apache-2.0 |
+
+---
+
+## DELPHI 5-Phase Cascade (signature inference mode)
+
+The VIDRAFT DELPHI cascade routes each question through 5 progressively deeper inference stages:
+
+1. **P1** — greedy single-shot (temperature 0)
+2. **P2** — MAJ@8 majority vote (temperature 0.7)
+3. **P3** — 16-vote tiebreak for close calls
+4. **P4** — Multi-Turn Inference (MTI): 3-turn self-critique × 8 chains
+5. **P5** — weighted global tiebreak across all phases
+
+Compute-optimal: most questions resolve at P1/P2; only ambiguous ones escalate.
+
+---
+
+## Usage
+
+### vLLM (recommended)
+
+```bash
+vllm serve FINAL-Bench/Darwin-218B-Delphi \
+    --tensor-parallel-size 8 \
+    --dtype bfloat16 \
+    --max-model-len 65536 \
+    --trust-remote-code \
+    --enforce-eager \
+    --limit-mm-per-prompt '{"image":0,"video":0}'
+```
+
+Requires vLLM ≥ 0.21.0 (`Cohere2VisionForConditionalGeneration` support).
+
+### Transformers
+
+```python
+from transformers import AutoModelForCausalLM, AutoTokenizer
+import torch
+
+model = AutoModelForCausalLM.from_pretrained(
+    "FINAL-Bench/Darwin-218B-Delphi",
+    dtype=torch.bfloat16,
+    device_map="auto",
+    trust_remote_code=True,
+)
+tok = AutoTokenizer.from_pretrained("FINAL-Bench/Darwin-218B-Delphi")
+
+messages = [
+    {"role": "user", "content": "Explain the SN2 mechanism step by step, "
+                                "then justify why CH3I reacts faster than CH3Cl."}
+]
+prompt = tok.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
+inputs = tok(prompt, return_tensors="pt").to(model.device)
+out = model.generate(**inputs, max_new_tokens=2048, temperature=0.3, top_p=0.9)
+print(tok.decode(out[0][inputs.input_ids.shape[1]:], skip_special_tokens=True))
+```
+
+---
+
+## License
+
+**Apache License 2.0**
+
+Built upon `CohereLabs/command-a-plus-05-2026-bf16` (Apache-2.0) and `Darwin-218B-kr` (Apache-2.0). All upstream components are permissively licensed.
+
+---
+
+## Contributors
+
+**Lead Architect & Developer** — 장재원 (Jaewon Jang), CTO, VIDRAFT
+*Domain SFT distillation pipeline, DELPHI cascade design, model integration.*
+
+**Organization** — VIDRAFT / FINAL-Bench
+https://huggingface.co/FINAL-Bench
+
+---
+
+## Citation
+
+```bibtex
+@misc{darwin-218b-delphi-2026,
+  title  = {Darwin-218B-Delphi: Chemistry-Specialized 218B MoE with DELPHI Cascade Inference},
+  author = {Jang, Jaewon and {VIDRAFT FINAL-Bench Team}},
+  year   = {2026},
+  publisher = {Hugging Face},
+  howpublished = {\url{https://huggingface.co/FINAL-Bench/Darwin-218B-Delphi}}
+}
+```