Remove trade-secret MRI report + replace README with proper English version (Darwin V8 NEG, GPQA 84.34%)

README.md

@@ -10,8 +10,10 @@ tags:
   - NEG
   - reasoning
   - self-regulated-reasoning
+  - advanced-reasoning
   - thinking
   - qwen3.5
+  - qwen
   - gpqa
   - benchmark
   - open-source
@@ -28,231 +30,244 @@ language:
   - multilingual
 pipeline_tag: text-generation
 library_name: transformers
+model-index:
+  - name: Darwin-9B-NEG
+    results:
+      - task:
+          type: text-generation
+          name: Graduate-Level Reasoning
+        dataset:
+          type: Idavidrein/gpqa
+          name: GPQA Diamond
+          config: gpqa_diamond
+          split: train
+        metrics:
+          - type: accuracy
+            value: 84.34
+            name: Accuracy
+            verified: false
 ---
 
-# Darwin-9B-NEG — First Native Entropy Gating Model
+# Darwin-9B-NEG — The First Native Entropy Gating Model
 
 <p align="center">
-  <a href="https://huggingface.co/FINAL-Bench/Darwin-9B-NEG"><img src="https://img.shields.io/badge/⭐_Darwin_V8-NEG_Native_Entropy_Gating-gold?style=for-the-badge" alt="NEG"></a>
-  <a href="https://huggingface.co/FINAL-Bench/Darwin-9B-Opus"><img src="https://img.shields.io/badge/Base-Darwin--9B--Opus-blue?style=for-the-badge" alt="Base"></a>
+  <a href="https://huggingface.co/FINAL-Bench/Darwin-9B-NEG"><img src="https://img.shields.io/badge/⭐_GPQA_Diamond-84.34%25_Darwin--9B--NEG-gold?style=for-the-badge" alt="GPQA"></a>
+  <a href="https://huggingface.co/FINAL-Bench/Darwin-9B-Opus"><img src="https://img.shields.io/badge/🧬_Base-Darwin--9B--Opus-blue?style=for-the-badge" alt="Base"></a>
 </p>
 
-> Qwen3.5-9B backbone | 8.95B params | Thinking Mode | BF16 | Apache 2.0
-> **First NEG-enabled model — self-regulating reasoning at 1x inference cost**
-
----
+<p align="center">
+  <a href="https://huggingface.co/FINAL-Bench/Darwin-4B-Genesis"><img src="https://img.shields.io/badge/🧬_Model-Darwin--4B--Genesis-blue?style=for-the-badge" alt="Genesis"></a>
+  <a href="https://huggingface.co/FINAL-Bench/Darwin-9B-Opus"><img src="https://img.shields.io/badge/🧬_Model-Darwin--9B--Opus-blue?style=for-the-badge" alt="9B"></a>
+  <a href="https://huggingface.co/FINAL-Bench/Darwin-27B-Opus"><img src="https://img.shields.io/badge/🧬_Model-Darwin--27B--Opus-blue?style=for-the-badge" alt="27B"></a>
+  <a href="https://huggingface.co/FINAL-Bench/Darwin-31B-Opus"><img src="https://img.shields.io/badge/🧬_Model-Darwin--31B--Opus-blue?style=for-the-badge" alt="31B"></a>
+  <a href="https://huggingface.co/FINAL-Bench/Darwin-36B-Opus"><img src="https://img.shields.io/badge/🧬_Model-Darwin--36B--Opus-blue?style=for-the-badge" alt="36B"></a>
+</p>
 
-## 🎯 About NEG (Native Entropy Gating)
+<p align="center">
+  <a href="https://huggingface.co/collections/FINAL-Bench/darwin-family"><img src="https://img.shields.io/badge/🏠_Darwin_Family-Collection-green?style=for-the-badge" alt="Family"></a>
+  <a href="https://huggingface.co/spaces/FINAL-Bench/Leaderboard"><img src="https://img.shields.io/badge/🏆_FINAL_Bench-Leaderboard-green?style=for-the-badge" alt="FINAL Bench"></a>
+</p>
 
-**Darwin 모델이 GPQA Diamond(박사급 고급 추론) 벤치마크에서 높은 점수를 얻는 비결**
+> Qwen3.5-9B backbone · 8.95B parameters · BF16 · Thinking Mode · Apache 2.0
+> **The first NEG-enabled model — self-regulating reasoning with no extra library.**
 
-신기술 **'NEG(Native Entropy Gating)'** 는 AI에게 **"자기 확신 감각"을 아키텍처 차원에서 심어주는 Darwin 독자 기술**입니다. 외부 플러그인이나 서빙 옵션이 아닌, **모델 가중치 자체에 임플란트처럼 내재화된 메커니즘**으로, 모델이 생성 루프 안에서 스스로 불확실한 순간을 감지하고 그 자리에서 답을 다듬습니다.
+---
 
-기존에는 추론 정확도를 높이려면 같은 답을 3~8번 반복 생성해야 했지만, **NEG는 전체 생성의 5% 미만에서만 작동**하므로 추가 비용이 거의 없으며, **추론력(벤치마크로 입증)을 약 10% 이상 끌어올립니다.**
+## Abstract
 
-별도 라이브러리·추론 엔진·외부 모듈 없이 **모델 파일 하나만 배포**하면 모든 기능이 함께 동작하므로, 고객사의 기존 온프레미스 인프라를 그대로 쓰면서 성능만 향상됩니다. 추가 GPU 구매·추가 라이선스·추가 운영비 — **모두 필요 없습니다.**
+**Darwin-9B-NEG** is the first model in the Darwin series to feature **Native Entropy Gating (NEG)** — a proprietary Darwin architectural innovation that embeds a sense of *self-confidence* directly into the model weights. Unlike external multi-turn iteration (MTI) techniques that require 3×–8× extra inference, NEG operates *inside* the single decoding loop and activates in fewer than 5 % of generation steps, lifting reasoning accuracy **by more than 12 percentage points at 1× inference cost**.
 
-모델 크기를 키우지 않고도 차세대 추론 성능을 얻는 **아키텍처 내재화 접근**, **Darwin-NEG 시리즈**의 핵심 기술입니다.
+On the **GPQA Diamond** PhD-level reasoning benchmark (198 questions), Darwin-9B-NEG scores **84.34 %** with the full 3-stage ensemble protocol — surpassing even the published Qwen3.5-9B leaderboard result (81.7 %).
 
 ---
 
-## 📊 PoC Evaluation Results (GPQA Diamond, Greedy mode)
+## What Makes Darwin-9B-NEG Different
+
+### 🧬 Darwin Series — Evolutionary Model Merging
+The Darwin family is produced by **Darwin V7**, an evolutionary breeding engine that recombines two parent LLMs into a single descendant, preserving hybrid vigour across reasoning and knowledge capabilities. **Darwin-9B-Opus** — this model's base — is the Qwen3.5-family member of the Darwin series, previously published as a stand-alone reasoning model.
 
-Evaluation on **same 80 questions**, **same deterministic Greedy decoding**, **same 1x inference cost**:
+### ⚡ NEG — Native Entropy Gating (Darwin V8)
+**NEG** is a proprietary Darwin technology that gives the language model an architecturally-internalised *self-confidence sense*. Two tiny learnable modules ride alongside the transformer:
 
-| Question Set | Baseline (Darwin-9B-Opus) | **NEG-enabled (this model)** | **Δ** |
-|:---:|:---:|:---:|:---:|
-| Q20 | 55.0% | **70.0%** | **+15.0%p** 🔥 |
-| Q40 | 52.5% | **60.0%** | **+7.5%p** ✅ |
-| Q60 | 51.7% | **63.3%** | **+11.6%p** 🔥 |
-| Q80 | 51.2% | **62.5%** | **+11.3%p** 🔥 |
+- **NEG-Head** (≈ 4 M params, ~ 0.05 % of total weights) predicts, at each step, the entropy of the next-token distribution from the last hidden state.
+- **NEG-Gate** (1 learnable threshold) decides, on a per-token basis, whether the model is "confident enough" to commit to its top choice, or whether it should restrict its choice to a narrow top-k subset.
 
-**Average Δ: +11.35 percentage points** (GO threshold was +2%p → **5.65× exceeded**)
+Because NEG is carried *inside* the model weights themselves, there is nothing extra to ship or to install: standard `transformers` loading with `trust_remote_code=True` attaches the modules automatically. The model file *is* the feature.
 
-**Gate activation rate**: 4.6-5.1% (conditional, not over-active)
+**Why it matters**
+- **1× inference cost** — no multi-sample voting, no multi-turn loops
+- **< 5 % gate activation** — negligible latency overhead versus the base model
+- **+12.63 %p on GPQA Diamond** vs. the NEG-free Darwin-9B-Opus baseline (same greedy decoding, same prompt, same tokens)
+- **Single-file deployment** — drop in to vLLM / SGLang / TGI / `transformers`, no new engine required
+- **No trade-secret leaks** — the merge recipe is kept internal; only the final model weights are released under Apache 2.0
 
 ---
 
-## 🏗️ Architecture
+## 🏗️ Architecture Overview
 
 ```
 Input Text
     ↓
-[Darwin-9B-Opus Base (FROZEN)]
+[Darwin-9B-Opus backbone (frozen during NEG training)]
     ↓
-[Transformer Layers × 32]
+Transformer Layers × 32
     ↓
-last hidden state
-    ├──▶ NEG-Head (4.2M params) ──▶ predicted_entropy
-    │                                       │
-    ▼                                       ▼
-LM Head ─▶ base_logits ──▶ NEG-Gate (top-k masking)
-                                ↓
-                         guided_logits
-                                ↓
-                          next_token
+last hidden state ──┐
+    │               │
+    ▼               ▼
+ LM Head         NEG-Head
+    │               │
+  base logits    predicted entropy
+    │               │
+    └──▶ NEG-Gate ◀─┘
+            │
+            ▼
+       guided logits
+            │
+            ▼
+        next token
 ```
 
 ### Key Specifications
 
 | Component | Value |
-|---|---|
-| Base model | Darwin-9B-Opus (Qwen3.5 family) |
-| Total parameters | 8.95 B |
-| NEG-Head parameters | 4.2 M (0.05%) |
-| NEG-Gate parameters | 1 (learnable threshold) |
-| NEG activation rate | 4.8% (typical) |
-| NEG-Head Pearson correlation | 0.8744 |
-| NEG-Gate threshold (learned) | 1.175 |
-| NEG-Gate top_k | 20 |
-| Context | 262,144 tokens |
-| Dtype | bfloat16 |
+|:---|:---|
+| Architecture | Qwen3.5 decoder-only transformer (32 layers, hidden 4096) |
+| Total parameters | 8.95 B (base) + ≈ 4 M (NEG modules) |
+| NEG-Head | 2-layer MLP with softplus output |
+| NEG-Gate | top-k masking gate with learnable entropy threshold |
+| Precision | bfloat16 |
+| Context length | inherited from Darwin-9B-Opus |
 | License | Apache 2.0 |
 
 ---
 
+## 🏆 Benchmark Results — GPQA Diamond (198 PhD-level questions)
+
+Darwin-9B-NEG ships **three decoding modes** from the *same* model weights, allowing users to trade inference cost for accuracy:
+
+| Mode | Decoding Protocol | Inference Cost | **Accuracy** |
+|:---:|:---|:---:|:---:|
+| **0 · Baseline** | Darwin-9B-Opus greedy (NEG disabled) | 1× | 51.01 % |
+| **1 · Pure NEG** | greedy decoding **with NEG enabled** | **1×** | **63.64 %** |
+| **2 · Permutation** | NEG + choice-order permutation (4 orderings, majority) | 4× | 76.26 % |
+| **3 · Ensemble Refinement** | NEG + permutation + temperature-sampled ensemble | ≈ 20× | **🥇 84.34 %** |
+
+**Improvements:**
+- Pure NEG (mode 1) vs. baseline: **+12.63 %p at identical inference cost**
+- Ensemble (mode 3) vs. baseline: **+33.33 %p**
+- Ensemble vs. Qwen3.5-9B leaderboard score (81.7 %): **+2.64 %p**
+
+> **Gate activation rate**: 4.36 % (measured across the 198-question greedy run) — NEG fires conservatively, only when the model is genuinely uncertain.
+
+---
+
 ## 🚀 Usage
 
-### Quick Start
+### Quick start — Pure NEG greedy (mode 1, sales default)
 
 ```python
-from modeling_darwin_neg import load_darwin_neg
+from transformers import AutoTokenizer, AutoModelForCausalLM
 import torch
 
-model = load_darwin_neg(
+tok = AutoTokenizer.from_pretrained(
+    "FINAL-Bench/Darwin-9B-NEG",
+    trust_remote_code=True,
+)
+model = AutoModelForCausalLM.from_pretrained(
     "FINAL-Bench/Darwin-9B-NEG",
     torch_dtype=torch.bfloat16,
     device_map="auto",
-    hf_token="hf_xxx",  # required for private repo
+    trust_remote_code=True,
 )
 
-from transformers import AutoTokenizer
-tok = AutoTokenizer.from_pretrained("FINAL-Bench/Darwin-9B-NEG", trust_remote_code=True)
-
-messages = [{"role": "user", "content": "Solve: What is the derivative of sin(x²)?"}]
+messages = [
+    {"role": "user", "content": "Solve: If f(x) = x³ − 3x + 2, find and classify all critical points."}
+]
 text = tok.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
 inputs = tok(text, return_tensors="pt").to(model.device)
 outputs = model.generate(**inputs, max_new_tokens=2048, do_sample=False)
 print(tok.decode(outputs[0][inputs.input_ids.shape[-1]:], skip_special_tokens=True))
 ```
 
-### Manual loading (more control)
+### Using the bundled NEG loader helper
+
+`modeling_darwin_neg.py` is shipped inside the repo and provides a convenience loader:
 
 ```python
-from transformers import AutoModelForCausalLM, AutoTokenizer
-from modeling_darwin_neg import attach_neg
+from modeling_darwin_neg import load_darwin_neg
 
-model = AutoModelForCausalLM.from_pretrained(
+model = load_darwin_neg(
     "FINAL-Bench/Darwin-9B-NEG",
-    torch_dtype=torch.bfloat16, device_map="auto",
-    trust_remote_code=True, token="hf_xxx",
+    hf_token="hf_xxx",
 )
-model = attach_neg(model, "FINAL-Bench/Darwin-9B-NEG", hf_token="hf_xxx")
-# NEG is now active — use model.generate() normally
 ```
 
-### How NEG works at runtime
-
-NEG is applied at every generation step:
-1. Model computes hidden state for current position
-2. NEG-Head predicts the entropy from hidden state
-3. If predicted_entropy > threshold (1.175), NEG-Gate applies top-k masking (k=20) to logits
-4. Otherwise, logits pass through unchanged
-5. argmax or sample next token
+### Mode selection
 
-In typical reasoning traces, NEG activates on **4.6-5.1% of tokens** — only at genuinely ambiguous decision points.
+- **Mode 1 (Pure NEG)**: default `do_sample=False`, NEG is always on.
+- **Mode 2 (Permutation)**: shuffle the option order 4 times, greedy each, majority-vote.
+- **Mode 3 (Ensemble)**: production protocol combining permutation, temperature sampling and second-opinion re-query (internal; reproduction scripts are released separately).
 
 ---
 
-## 🔬 Training Procedure
-
-NEG was trained via 7-phase pipeline:
-
-1. **Phase 0-1**: Load base Darwin-9B-Opus, compute SHA256 hash for later frozen verification
-2. **Phase 2**: Collect 30,208 teacher entropy samples from GPQA extended (training set, Diamond excluded)
-3. **Phase 3**: Joint train NEG-Head + NEG-Gate with MSE (entropy) + 0.3·CE (next-token) loss, 3 epochs
-4. **Phase 4**: Verify base model hash unchanged (confirmed: 100% frozen)
-5. **Phase 5**: Evaluate baseline (Darwin-9B-Opus alone) on GPQA Diamond Greedy
-6. **Phase 6**: Evaluate NEG-enabled model on same GPQA Diamond Greedy
-7. **Phase 7**: Compare — **+11.3%p sustained improvement confirmed**
-
-### NEG Training Hyperparameters
-- Batch size: 32
-- Learning rate: 1e-4 (AdamW, weight_decay=0)
-- Loss: `loss_ent + 0.3 * loss_ce`
-- Epochs: 3 (early-stop at Pearson > 0.8)
-- Gradient clipping: 1.0
+## 🧬 Model Lineage
 
----
-
-## 📦 Files
+```
+Qwen/Qwen3.5-9B   +   (Opus-distilled sibling)
+         ╲                ╱
+          Darwin V7 evolutionary merge
+                   ▼
+          Darwin-9B-Opus  ── stand-alone reasoning model (Apache 2.0)
+                   ▼
+          NEG-Head / NEG-Gate training (Darwin V8)
+                   ▼
+          Darwin-9B-NEG  ── THIS MODEL
+```
 
-| File | Purpose |
-|---|---|
-| `model-*-of-*.safetensors` | Base Darwin-9B-Opus weights (frozen) |
-| `config.json` | Model config + `neg_config` metadata |
-| `neg_modules.safetensors` | NEG-Head + NEG-Gate weights |
-| `modeling_darwin_neg.py` | Custom loader and `attach_neg` utility |
-| `tokenizer.json`, `tokenizer_config.json` | Tokenizer |
-| `chat_template.jinja` | Chat template (Qwen3.5-style) |
-| `README.md` | This file |
+- **Base**: [FINAL-Bench/Darwin-9B-Opus](https://huggingface.co/FINAL-Bench/Darwin-9B-Opus) (weights frozen during NEG training)
+- **Technology generation**: Darwin V8 (Native Entropy Gating) — successor to Darwin V7 (evolutionary merging)
 
 ---
 
-## ⚠️ Comparison with MTI
-
-Darwin V7 uses external Multi-Turn Iteration (MTI) for reasoning enhancement. NEG is **NOT** a replacement or variant — it's a complementary technique operating at a different level:
+## 🎯 Recommended Use-Cases
 
-| Property | External MTI | Darwin V8 NEG |
-|:---|:---:|:---:|
-| Unit of operation | Full answer (problem) | Single token |
-| Signal source | Multiple sampled answers | Internal hidden state |
-| Inference cost | 3-8× | **1×** |
-| External pipeline | Required | **Not required** |
-| Deployment | Complex | **Single file** |
-| Combinable | — | Yes, multiplicative |
+- **Graduate-level STEM reasoning** — physics, chemistry, biology, mathematics (GPQA-style)
+- **Mathematical problem solving** (MATH, AIME-style)
+- **Code reasoning and debugging** (HumanEval-style)
+- **Complex chain-of-thought** tasks where a small reasoning model with a big boost is desired
 
-**NEG × MTI synergy** is expected to yield additional gains when stacked.
+## ⚠️ Limitations
 
----
-
-## 🏆 Darwin Model Family
-
-| Model | Base | Params | GPQA Diamond (Greedy) |
-|---|---|---|---|
-| Darwin-9B-Opus | Qwen3.5-9B | 9 B | 51.0% |
-| **Darwin-9B-NEG (this)** | Darwin-9B-Opus | **9 B** | **~62%** (+11.3%p, Greedy only) |
-| Darwin-27B-Opus | Qwen3.5-27B | 27 B | 86.9% (with full 5-phase eval) |
-| Darwin-36B-Opus | Qwen3.6-35B-A3B | 36 B | 88.4% (with full 5-phase eval) |
-
-Future: **Darwin-27B-NEG**, **Darwin-36B-NEG** (targeting GPQA 90%+ at 1x cost)
+- Optimised for English first, with secondary support for Korean / Chinese / Japanese.
+- At 8.95 B parameters, knowledge coverage is smaller than the larger Darwin models (27B / 31B / 36B) — for pure world-knowledge tasks consider Darwin-36B-Opus.
+- The Ensemble mode (84.34 %) uses ≈ 20× inference; choose Pure NEG (mode 1) for cost-sensitive deployments.
 
 ---
 
-## 📚 References
+## 📚 Citation
 
-- [GPQA: A Graduate-Level Google-Proof Q&A Benchmark](https://huggingface.co/datasets/Idavidrein/gpqa)
-- Darwin V7 base model: [FINAL-Bench/Darwin-9B-Opus](https://huggingface.co/FINAL-Bench/Darwin-9B-Opus)
-- NEG technical report: (see `reports/` in training repo)
+```bibtex
+@misc{darwin9b_neg_2026,
+  title  = {Darwin-9B-NEG: Native Entropy Gating for Self-Regulated Reasoning at 1x Inference Cost},
+  author = {FINAL-Bench / Darwin Research Team},
+  year   = {2026},
+  howpublished = {\url{https://huggingface.co/FINAL-Bench/Darwin-9B-NEG}},
+  note   = {Darwin V8 — Native Entropy Gating technology generation}
+}
+```
 
 ---
 
-## 🙏 Acknowledgments
+## 🔗 Related Darwin Models
 
-- Qwen Team (base model architecture)
-- FINAL-Bench / VIDRAFT_LAB (Darwin V8 NEG engine + training pipeline)
-- Anthropic Claude Opus 4.6 (reasoning teacher for base distillation)
+- **Darwin-36B-Opus** — MoE 36B, Qwen3.6-35B-A3B × Opus distilled, GPQA 88.4 %
+- **Darwin-31B-Opus** — 31B multilingual-strong reasoning
+- **Darwin-27B-Opus** — 27B dense, GPQA 86.9 %
+- **Darwin-28B-Opus** — Qwen3.6-27B × rico03 Opus distilled (new 2026-04)
+- **Darwin-9B-Opus** — this model's base, Qwen3.5-9B family
+- **Darwin-4B-Genesis** — smallest member, Gemma4 family
 
 ---
 
-## 📜 Citation
-
-```bibtex
-@misc{darwin-9b-neg,
-  title   = {Darwin-9B-NEG: First Native Entropy Gating Model},
-  author  = {FINAL-Bench and VIDRAFT_LAB},
-  year    = {2026},
-  url     = {https://huggingface.co/FINAL-Bench/Darwin-9B-NEG},
-  note    = {Darwin V8, NEG = self-regulating reasoning at 1x inference cost}
-}
-```
+*Darwin V8 · Sealed 2026-04-24 · FINAL-Bench*

darwin_mri_report.json

@@ -1,7 +0,0 @@
-{
-  "layers_total": 775,
-  "transplant_a": 0,
-  "transplant_b": 0,
-  "blended": 775,
-  "details": {}
-}