Model card: remove adapter/PEFT framing -> full standalone model (AutoModel usage, base_model_relation=finetune)

README.md

@@ -21,9 +21,6 @@ tags:
   - test-time-compute
   - qwen3.6
   - qwen
-  - lora
-  - peft
-  - adapter
   - gpqa
   - benchmark
   - open-source
@@ -33,7 +30,7 @@ tags:
   - eval-results
 base_model:
   - FINAL-Bench/Darwin-28B-Opus
-base_model_relation: adapter
+base_model_relation: finetune
 model-index:
   - name: Darwin-28B-REASON
     results:
@@ -70,16 +67,16 @@ model-index:
   <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>
 
-> Reasoning-enhanced model built on Darwin-28B-Opus · Reasoning-Trace Distillation (RTD) · Darwin-DELPHI test-time engine · BF16 · Apache 2.0
+> Full standalone reasoning model derived from Darwin-28B-Opus · Reasoning-Trace Distillation (RTD) · Darwin-DELPHI test-time engine · 27.6 B · BF16 · Apache 2.0
 > **GPQA Diamond: 89.39 % with Darwin-DELPHI**
 
 ---
 
 ## Overview
 
-**Darwin-28B-REASON** is a reasoning-enhanced model built on top of **[Darwin-28B-Opus](https://huggingface.co/FINAL-Bench/Darwin-28B-Opus)**. It combines two components:
+**Darwin-28B-REASON** is a reasoning-enhanced **standalone model** derived from **[Darwin-28B-Opus](https://huggingface.co/FINAL-Bench/Darwin-28B-Opus)**. It combines two components:
 
-1. **Reasoning-Trace Distillation (RTD)** — a reasoning-trace distillation stage applied to the Darwin-28B-Opus base, delivered as a lightweight adapter.
+1. **Reasoning-Trace Distillation (RTD)** — a reasoning-trace distillation stage applied on top of the Darwin-28B-Opus base, producing this fully self-contained model (full weights, no external adapter required).
 2. **Darwin-DELPHI** — a proprietary test-time reasoning engine.
 
 Together they push graduate-level scientific reasoning to the top tier of the Darwin family: **89.39 %** on GPQA Diamond with Darwin-DELPHI. The model is released under **Apache-2.0**.
@@ -101,9 +98,9 @@ Together they push graduate-level scientific reasoning to the top tier of the Da
 | Role | Model | Contribution |
 |:---:|:---|:---|
 | **Base** | [`FINAL-Bench/Darwin-28B-Opus`](https://huggingface.co/FINAL-Bench/Darwin-28B-Opus) | GPQA #3 (88.89 %) Qwen3.6-generation reasoning backbone. |
-| **RTD adapter** | reasoning-trace distillation | Distills complete reasoning chains into a lightweight adapter on the Opus base. |
+| **RTD training** | reasoning-trace distillation | Distills complete reasoning chains into the model on top of the Opus base. |
 | **Test-time engine** | Darwin-DELPHI | Proprietary inference-time consensus engine (not stored in weights). |
-| **Result** | **`Darwin-28B-REASON`** (this model) | RTD adapter + Darwin-DELPHI → **89.39 %** GPQA Diamond. |
+| **Result** | **`Darwin-28B-REASON`** (this model) | Full standalone RTD model + Darwin-DELPHI → **89.39 %** GPQA Diamond. |
 
 ---
 
@@ -111,11 +108,13 @@ Together they push graduate-level scientific reasoning to the top tier of the Da
 
 | Component | Value |
 |:---|:---|
-| Base architecture | `Qwen3_5ForConditionalGeneration` (Qwen3.6 generation, hybrid linear + full attention) |
-| Base model | FINAL-Bench/Darwin-28B-Opus (27.6 B, BF16) |
-| Delivery | LoRA / PEFT adapter on the Darwin-28B-Opus base |
+| Architecture | `Qwen3_5ForConditionalGeneration` (Qwen3.6 generation, hybrid linear + full attention; text path, `language_model_only`) |
+| Parameters | **27.6 B** (BF16) — full standalone weights |
+| Layers | 64 (3 linear : 1 full attention, `full_attention_interval = 4`) |
+| Vocab size | 248 320 |
+| Context length | 262 144 (long-chain reasoning supported) |
+| Delivery | Full self-contained model — no external base or adapter required |
 | Precision | bfloat16 |
-| Context length | Inherited from base (long-chain reasoning supported) |
 | License | Apache 2.0 |
 
 ---
@@ -124,7 +123,7 @@ Together they push graduate-level scientific reasoning to the top tier of the Da
 
 ### ① RTD — Reasoning-Trace Distillation
 
-RTD distills **complete reasoning chains** from a publicly available mathematical corpus (Apache-2.0 source) into a lightweight adapter on the Darwin-28B-Opus base. The adapter strengthens long-form, multi-step scientific reasoning while preserving the base model's bilingual capability.
+RTD distills **complete reasoning chains** from a publicly available mathematical corpus (Apache-2.0 source) on top of the Darwin-28B-Opus base, producing this standalone model. It strengthens long-form, multi-step scientific reasoning while preserving the base model's bilingual capability.
 
 > The full RTD recipe (curation, trace selection, training schedule) is **proprietary** and is not disclosed.
 
@@ -151,25 +150,21 @@ The evaluation methodology for the Darwin-DELPHI result is **protected**; sample
 
 ## 🚀 Usage
 
-Darwin-28B-REASON ships as a **LoRA / PEFT adapter** on the Darwin-28B-Opus base.
+Darwin-28B-REASON is a **full standalone model** — load it directly, no base model or adapter merge required.
 
 ```python
 from transformers import AutoTokenizer, AutoModelForCausalLM
-from peft import PeftModel
 import torch
 
-BASE = "FINAL-Bench/Darwin-28B-Opus"
-ADAPTER = "FINAL-Bench/Darwin-28B-REASON"
+MODEL = "FINAL-Bench/Darwin-28B-REASON"
 
-tok = AutoTokenizer.from_pretrained(BASE, trust_remote_code=True)
-
-base = AutoModelForCausalLM.from_pretrained(
-    BASE,
+tok = AutoTokenizer.from_pretrained(MODEL, trust_remote_code=True)
+model = AutoModelForCausalLM.from_pretrained(
+    MODEL,
     torch_dtype=torch.bfloat16,
     device_map="auto",
     trust_remote_code=True,
 )
-model = PeftModel.from_pretrained(base, ADAPTER)
 model.eval()
 
 messages = [
@@ -182,7 +177,7 @@ outputs = model.generate(**inputs, max_new_tokens=2048)
 print(tok.decode(outputs[0][inputs.input_ids.shape[-1]:], skip_special_tokens=True))
 ```
 
-> The 89.39 % GPQA Diamond result is produced with the Darwin-DELPHI test-time engine, which is applied on top of this adapter. Darwin-DELPHI is provided through the Darwin-series evaluation harness.
+> The 89.39 % GPQA Diamond result is produced with the Darwin-DELPHI test-time engine applied on top of this model. Darwin-DELPHI is provided through the Darwin-series evaluation harness.
 
 ---
 
@@ -196,8 +191,8 @@ print(tok.decode(outputs[0][inputs.input_ids.shape[-1]:], skip_special_tokens=Tr
 
 ## ⚠️ Limitations
 
-- Requires the Darwin-28B-Opus base (≈ 55 GB VRAM in bfloat16) plus the adapter; a single A100-80GB or B200 is sufficient.
-- The 89.39 % result depends on the Darwin-DELPHI test-time engine; the adapter alone provides strong but lower single-model accuracy.
+- The 27.6 B model in bfloat16 requires ≈ 55 GB of VRAM (a single A100-80GB or B200 is sufficient).
+- The 89.39 % result depends on the Darwin-DELPHI test-time engine; the model on its own delivers strong but lower single-model accuracy.
 - Optimised for English first, with secondary support for Korean, Chinese, and Japanese.
 - Reasoning traces tend to be verbose — control with `max_new_tokens` as needed.
 
@@ -211,7 +206,7 @@ print(tok.decode(outputs[0][inputs.input_ids.shape[-1]:], skip_special_tokens=Tr
   author = {FINAL-Bench / Darwin Research Team},
   year   = {2026},
   howpublished = {\url{https://huggingface.co/FINAL-Bench/Darwin-28B-REASON}},
-  note   = {RTD adapter + Darwin-DELPHI · 89.39 % GPQA Diamond}
+  note   = {RTD + Darwin-DELPHI · 89.39 % GPQA Diamond}
 }
 
 @misc{darwin_family_2026,