Text Generation
Transformers
Safetensors
qwen3_5_text
darwin
darwin-v8
opus-distilled
qwen3.5
reasoning
korean
claude-opus
lora-merged
conversational
Instructions to use FINAL-Bench/Darwin-2B-Opus with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Libraries
- Transformers
How to use FINAL-Bench/Darwin-2B-Opus with Transformers:
# Use a pipeline as a high-level helper from transformers import pipeline pipe = pipeline("text-generation", model="FINAL-Bench/Darwin-2B-Opus") messages = [ {"role": "user", "content": "Who are you?"}, ] pipe(messages)# Load model directly from transformers import AutoTokenizer, AutoModelForCausalLM tokenizer = AutoTokenizer.from_pretrained("FINAL-Bench/Darwin-2B-Opus") model = AutoModelForCausalLM.from_pretrained("FINAL-Bench/Darwin-2B-Opus") messages = [ {"role": "user", "content": "Who are you?"}, ] inputs = tokenizer.apply_chat_template( messages, add_generation_prompt=True, tokenize=True, return_dict=True, return_tensors="pt", ).to(model.device) outputs = model.generate(**inputs, max_new_tokens=40) print(tokenizer.decode(outputs[0][inputs["input_ids"].shape[-1]:])) - Notebooks
- Google Colab
- Kaggle
- Local Apps Settings
- vLLM
How to use FINAL-Bench/Darwin-2B-Opus with vLLM:
Install from pip and serve model
# Install vLLM from pip: pip install vllm # Start the vLLM server: vllm serve "FINAL-Bench/Darwin-2B-Opus" # Call the server using curl (OpenAI-compatible API): curl -X POST "http://localhost:8000/v1/chat/completions" \ -H "Content-Type: application/json" \ --data '{ "model": "FINAL-Bench/Darwin-2B-Opus", "messages": [ { "role": "user", "content": "What is the capital of France?" } ] }'Use Docker
docker model run hf.co/FINAL-Bench/Darwin-2B-Opus
- SGLang
How to use FINAL-Bench/Darwin-2B-Opus with SGLang:
Install from pip and serve model
# Install SGLang from pip: pip install sglang # Start the SGLang server: python3 -m sglang.launch_server \ --model-path "FINAL-Bench/Darwin-2B-Opus" \ --host 0.0.0.0 \ --port 30000 # Call the server using curl (OpenAI-compatible API): curl -X POST "http://localhost:30000/v1/chat/completions" \ -H "Content-Type: application/json" \ --data '{ "model": "FINAL-Bench/Darwin-2B-Opus", "messages": [ { "role": "user", "content": "What is the capital of France?" } ] }'Use Docker images
docker run --gpus all \ --shm-size 32g \ -p 30000:30000 \ -v ~/.cache/huggingface:/root/.cache/huggingface \ --env "HF_TOKEN=<secret>" \ --ipc=host \ lmsysorg/sglang:latest \ python3 -m sglang.launch_server \ --model-path "FINAL-Bench/Darwin-2B-Opus" \ --host 0.0.0.0 \ --port 30000 # Call the server using curl (OpenAI-compatible API): curl -X POST "http://localhost:30000/v1/chat/completions" \ -H "Content-Type: application/json" \ --data '{ "model": "FINAL-Bench/Darwin-2B-Opus", "messages": [ { "role": "user", "content": "What is the capital of France?" } ] }' - Docker Model Runner
How to use FINAL-Bench/Darwin-2B-Opus with Docker Model Runner:
docker model run hf.co/FINAL-Bench/Darwin-2B-Opus
| license: apache-2.0 | |
| base_model: Qwen/Qwen3.5-2B | |
| tags: | |
| - darwin | |
| - darwin-v8 | |
| - opus-distilled | |
| - qwen3.5 | |
| - reasoning | |
| - korean | |
| - claude-opus | |
| - lora-merged | |
| language: | |
| - en | |
| - ko | |
| - zh | |
| - ja | |
| pipeline_tag: text-generation | |
| library_name: transformers | |
| # ๐ง Darwin-2B-Opus | |
| **Darwin V8 ์๋ฆฌ์ฆ์ 2B ๊ฒฝ๋ ๋ชจ๋ธ** | |
| Claude Opus 4.5/4.6 ๋ฐ Sonnet 4.6์ ์ถ๋ก ์คํ์ผ์ ์ฃผ์ ํ Qwen3.5-2B ๊ธฐ๋ฐ ๋ชจ๋ธ. | |
| --- | |
| ## ๐งฌ ๊ฐ๊ณ๋ (Pedigree) | |
| - ๐จ **Father (Base)**: [`Qwen/Qwen3.5-2B`](https://huggingface.co/Qwen/Qwen3.5-2B) | |
| - ๐ฉ **Mother (LoRA Adapter)**: [`FINAL-Bench/Darwin-2B-Opus-LoRA`](https://huggingface.co/FINAL-Bench/Darwin-2B-Opus-LoRA) | |
| - ๐ถ **Child (This model)**: `FINAL-Bench/Darwin-2B-Opus` โ merged full-weight standalone | |
| --- | |
| ## ๐ Darwin V8 ์๋ฆฌ์ฆ ์ ๋ณด | |
| | ํญ๋ชฉ | ๊ฐ | | |
| |------|-----| | |
| | ๋ชจ๋ธ ํฌ๊ธฐ | 2.3B ํ๋ผ๋ฏธํฐ | | |
| | ์ํคํ ์ฒ | Qwen3.5 (hybrid attention) | | |
| | ํ์ต ๋ฐฉ์ | SFT with LoRA (all-linear, rank=16) | | |
| | ํ์ต ๋ฐ์ดํฐ | 9,762 ์ํ (Claude Opus/Sonnet + ํ๊ตญ์ด reasoning) | | |
| | ํ์ต ์๊ฐ | 29๋ถ (8รB200 GPU) | | |
| | ์ต์ข Loss | 0.837 | | |
| | Token Accuracy | 76.6% | | |
| ### ๐ ๋ฒค์น๋งํฌ (GPQA Diamond 198) | |
| - **์ ํ๋**: 37.37% (74/198) | |
| - **๋ต๋ณ ์ถ์ถ ์ฑ๊ณต๋ฅ ๊ธฐ์ค ์ ๋ต๋ฅ **: 50.7% | |
| --- | |
| ## ๐ ๋น ๋ฅธ ์ฌ์ฉ๋ฒ | |
| ```python | |
| from transformers import AutoTokenizer, AutoModelForCausalLM | |
| import torch | |
| model_id = "FINAL-Bench/Darwin-2B-Opus" | |
| tok = AutoTokenizer.from_pretrained(model_id, trust_remote_code=True) | |
| model = AutoModelForCausalLM.from_pretrained( | |
| model_id, torch_dtype=torch.bfloat16, device_map="auto", trust_remote_code=True | |
| ) | |
| messages = [ | |
| {"role": "user", "content": "2024๋ ํ๊ตญ ์ต์ ์๊ธ 9,860์์ด๋ค. ์ฃผ 40์๊ฐ ร 4์ฃผ ์๊ธ์?"} | |
| ] | |
| prompt = tok.apply_chat_template(messages, tokenize=False, add_generation_prompt=True) | |
| inputs = tok(prompt, return_tensors="pt").to(model.device) | |
| with torch.no_grad(): | |
| outputs = model.generate( | |
| **inputs, | |
| max_new_tokens=800, | |
| do_sample=False, | |
| pad_token_id=tok.eos_token_id, | |
| ) | |
| print(tok.decode(outputs[0][inputs.input_ids.shape[1]:], skip_special_tokens=True)) | |
| ``` | |
| --- | |
| ## ๐งฌ Darwin V8 ํ์ต ํ์ดํ๋ผ์ธ | |
| ``` | |
| [Qwen/Qwen3.5-2B] โโโโ Base ๋ชจ๋ธ (๋๊ฒฐ) | |
| + | |
| [9,762 Claude Opus/Sonnet + ํ๊ตญ์ด Reasoning ์ํ] | |
| โ | |
| [SFT Training] | |
| - LoRA (all-linear, r=16, ฮฑ=32) | |
| - Learning rate: 2e-4 (V8 rule: ร10 FullFT) | |
| - 2 epochs, bf16, 8รB200 DDP | |
| - Loss: 0.991 โ 0.837 (-15%) | |
| - Token accuracy: 73.9% โ 76.6% (+2.7%p) | |
| โ | |
| [LoRA merge into base weights] | |
| โ | |
| [Darwin-2B-Opus] โ ์ด ๋ชจ๋ธ | |
| ``` | |
| --- | |
| ## ๐ ํ์ต ๋ฐ์ดํฐ ๊ตฌ์ฑ | |
| | ์นดํ ๊ณ ๋ฆฌ | ์ํ ์ | % | ์ถ์ฒ | | |
| |---------|--------|---|-----| | |
| | General Reasoning | 4,422 | 45% | Opus 4.5/4.6, Sonnet 4.6 | | |
| | Math (English) | 1,960 | 20% | DeepSeek-v3.2 OpenR1-Math | | |
| | Code (English) | 1,680 | 17% | DeepSeek-v3.2 CodeReasoning + GPT-5 Codex | | |
| | Korean Thinking | 200 | 2% | Multilingual-Thinking-Korean | | |
| | **Korean Math** | **1,500** | **15%** | orca-math-word-problems-korean | | |
| | **ํฉ๊ณ (ํํฐ ํ)** | **9,762** | 100% | - | | |
| --- | |
| ## ๐ฏ Darwin V8 ์ค๊ณ ์ฒ ํ | |
| 1. **LoRA Without Regret** โ `all-linear` target, LR ร 10, rank=16์ผ๋ก ์ถฉ๋ถ | |
| 2. **Response Distillation** โ Pre-generated Opus traces๋ก ๋น์ฉ ํจ์จ์ ์ฆ๋ฅ | |
| 3. **ํ๊ตญ์ด Reasoning ๊ฐํ** โ KoAlpaca ๊ฐ๋จ QA ๋์ Claude ์ถ๋ก ๊ถค์ ์ฌ์ฉ | |
| 4. **Merge-and-Deploy** โ LoRA ์ด๋ํฐ ํตํฉ ํ ์ถ๊ฐ ์์กด์ฑ ์์ด ๋ฐฐํฌ | |
| --- | |
| ## ๐ ์ํ ํ ์คํธ ๊ฒฐ๊ณผ (5๋ฌธ์ ) | |
| | ์ ํ | ์ ๋ต | ๋น๊ณ | | |
| |-----|:---:|-----| | |
| | ์์ด ์ํ (๊ธฐ์ฐจ ์๋) | โ 80 km/h | LaTeX ๋จ๊ณ๋ณ ํ์ด | | |
| | ์์ด ๋ ผ๋ฆฌ (ํค ๋น๊ต) | โ Carol | ์ถ์ด์จ ๋ช ์ | | |
| | ์์ด ์ฝ๋ (์์ ํ๋ณ) | โ ์ ํ | docstring + ๋ณต์ก๋ ๋ถ์ | | |
| | **ํ๊ตญ์ด ์๊ธ ๊ณ์ฐ** | โ **1,577,600์** | ๋จ๊ณ๋ณ ํ๊ตญ์ด ์ค๋ช | | |
| | **ํ๊ตญ์ด ์ฐ๋ฆฝ๋ฐฉ์ ์** | โ **1,200์** | ์ ์ ํ์ด + ๊ฒ์ฆ | | |
| **5/5 ์ ๋ต** โ ์์ด+ํ๊ตญ์ด ๋ชจ๋ ์๋ฒฝ โญ | |
| --- | |
| ## โ ๏ธ ์ ํ ์ฌํญ | |
| - **๊ท๋ชจ**: 2.3B ํ๋ผ๋ฏธํฐ (Darwin ์๋ฆฌ์ฆ ์ต์) | |
| - **GPQA Diamond**: 37.37% (๋ํ ๋ชจ๋ธ ๋๋น ๋ฎ์ง๋ง 2B ์ค ์ต๊ณ ์์ค) | |
| - **๊ธด ์ปจํ ์คํธ**: ํ์ต ์ `max_length=4,096`๋ก ํ์ต๋จ | |
| - **์ง์ ํ๊ณ**: 2B ๋ชจ๋ธ์ ๋ฐฑ๊ณผ์ฌ์ ์ ์ง์ ํ๊ณ ์์ | |
| --- | |
| ## ๐ ๊ด๋ จ ๋ชจ๋ธ | |
| - ๐งฉ [`FINAL-Bench/Darwin-2B-Opus-LoRA`](https://huggingface.co/FINAL-Bench/Darwin-2B-Opus-LoRA) โ ์ด ๋ชจ๋ธ์ **LoRA ์ด๋ํฐ ๋จ๋ ๋ฒ์ ** (67MB) | |
| - โก [`FINAL-Bench/Darwin-2B-Opus-ONNX`](https://huggingface.co/FINAL-Bench/Darwin-2B-Opus-ONNX) โ **๋ธ๋ผ์ฐ์ /WebGPU์ฉ ONNX ์์ํ ๋ฒ์ ** (์์ ) | |
| ### ๐ Darwin ์๋ฆฌ์ฆ | |
| - [`Darwin-31B-Opus`](https://huggingface.co/FINAL-Bench/Darwin-31B-Opus) โ GPQA 85.9% | |
| - [`Darwin-27B-Opus`](https://huggingface.co/FINAL-Bench/Darwin-27B-Opus) โ GPQA 86.9% | |
| - [`Darwin-9B-Opus`](https://huggingface.co/FINAL-Bench/Darwin-9B-Opus) | |
| - [`Darwin-4B-Opus`](https://huggingface.co/FINAL-Bench/Darwin-4B-Opus) | |
| - **Darwin-2B-Opus** (์ด ๋ชจ๋ธ) โญ ์ต๊ฒฝ๋ | |
| --- | |
| ## ๐ชช ๋ผ์ด์ ์ค | |
| - Base model: Apache 2.0 (Qwen) | |
| - ํ์ต ๋ฐ์ดํฐ: ๊ฐ ๋ฐ์ดํฐ์ ๊ฐ๋ณ ๋ผ์ด์ ์ค ์ฐธ์กฐ | |
| - ์ด ๋ชจ๋ธ: Apache 2.0 | |
| --- | |
| ## ๐ ํฌ๋ ๋ง | |
| - **Base**: Qwen team (Alibaba) | |
| - **Teacher**: Anthropic (Claude Opus 4.5/4.6, Sonnet 4.6) | |
| - **๋ฐ์ดํฐ ๊ณต๊ฐ**: nohurry, TeichAI, kuotient, PoSTMEDIA | |
| - **Training & Release**: **FINAL-Bench / VIDRAFT_LAB** | |
| --- | |
| *Darwin V8 ยท Part of the evolutionary model series by FINAL-Bench* | |
| This model is introduced in [Darwin Family](https://arxiv.org/abs/2605.14386). |