Text Generation
Transformers
Safetensors
Korean
English
qwen3
chimera
darwin
vidraft
korean
kmmlu
reasoning
conversational
text-generation-inference
Instructions to use FINAL-Bench/Darwin-4B-Chimera with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Libraries
- Transformers
How to use FINAL-Bench/Darwin-4B-Chimera with Transformers:
# Use a pipeline as a high-level helper from transformers import pipeline pipe = pipeline("text-generation", model="FINAL-Bench/Darwin-4B-Chimera") messages = [ {"role": "user", "content": "Who are you?"}, ] pipe(messages)# Load model directly from transformers import AutoTokenizer, AutoModelForCausalLM tokenizer = AutoTokenizer.from_pretrained("FINAL-Bench/Darwin-4B-Chimera") model = AutoModelForCausalLM.from_pretrained("FINAL-Bench/Darwin-4B-Chimera") 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-4B-Chimera with vLLM:
Install from pip and serve model
# Install vLLM from pip: pip install vllm # Start the vLLM server: vllm serve "FINAL-Bench/Darwin-4B-Chimera" # 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-4B-Chimera", "messages": [ { "role": "user", "content": "What is the capital of France?" } ] }'Use Docker
docker model run hf.co/FINAL-Bench/Darwin-4B-Chimera
- SGLang
How to use FINAL-Bench/Darwin-4B-Chimera 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-4B-Chimera" \ --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-4B-Chimera", "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-4B-Chimera" \ --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-4B-Chimera", "messages": [ { "role": "user", "content": "What is the capital of France?" } ] }' - Docker Model Runner
How to use FINAL-Bench/Darwin-4B-Chimera with Docker Model Runner:
docker model run hf.co/FINAL-Bench/Darwin-4B-Chimera
Rewrite card: Chimera positioning + honest scope; remove proprietary method details and method tags
Browse files
README.md
CHANGED
|
@@ -1,22 +1,97 @@
|
|
| 1 |
---
|
| 2 |
license: gemma
|
| 3 |
base_model: FINAL-Bench/Darwin-V9-Chimera-4B-SFT
|
| 4 |
-
tags:
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 5 |
pipeline_tag: text-generation
|
|
|
|
|
|
|
|
|
|
|
|
|
| 6 |
---
|
| 7 |
|
| 8 |
-
# Darwin-
|
| 9 |
|
| 10 |
-
|
| 11 |
|
| 12 |
-
|
| 13 |
-
Self-rollout (n=8, temp 0.8) on KMMLU-domain train problems; keep correct CoT **only from learnable-edge questions** (1-6 of 8 correct), excluding already-stable (8/8) and unknown (0/8). Round 2 (STaR) rolls out from the improved model on fresh problems and accumulates edge data. LoRA r16, LR 2e-5, merged.
|
| 14 |
|
| 15 |
-
|
| 16 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 17 |
|---|---|---|
|
| 18 |
-
|
|
| 19 |
-
|
|
| 20 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 21 |
|
| 22 |
-
|
|
|
|
| 1 |
---
|
| 2 |
license: gemma
|
| 3 |
base_model: FINAL-Bench/Darwin-V9-Chimera-4B-SFT
|
| 4 |
+
tags:
|
| 5 |
+
- chimera
|
| 6 |
+
- darwin
|
| 7 |
+
- vidraft
|
| 8 |
+
- korean
|
| 9 |
+
- kmmlu
|
| 10 |
+
- reasoning
|
| 11 |
pipeline_tag: text-generation
|
| 12 |
+
language:
|
| 13 |
+
- ko
|
| 14 |
+
- en
|
| 15 |
+
library_name: transformers
|
| 16 |
---
|
| 17 |
|
| 18 |
+
# Darwin-4B-Chimera
|
| 19 |
|
| 20 |
+
**๋น๋๋ํํธ(VIDRAFT) Chimera ๊ธฐ์ ๋ก ๋ง๋ 4B ํ๊ตญ์ด ์ถ๋ก ๋ชจ๋ธ.**
|
| 21 |
|
| 22 |
+
์์ ๋ชจ๋ธ์ด ์๊ธฐ ํฌ๊ธฐ๋ฅผ ๋์ด์๊ฒ ๋ง๋๋ ๊ฒ์ด ๋ชฉํ์
๋๋ค. ํ๋ผ๋ฏธํฐ๋ฅผ ๋๋ฆฌ๋ ๋์ , ์๋ก ๋ค๋ฅธ ๊ฐ์ ์ **ํฉ์ฒด**ํ๊ณ ์ค์ค๋ก **์ ์ **ํ๊ฒ ํ์ต๋๋ค.
|
|
|
|
| 23 |
|
| 24 |
+
---
|
| 25 |
+
|
| 26 |
+
## Chimera๋
|
| 27 |
+
|
| 28 |
+
**Chimera**๋ ์๋ก ๋ค๋ฅธ ๊ณ์ดยท๊ฐ์ ์ ๊ฐ์ง ๋ชจ๋ธ์ ๊ตฌ์ฑ์์๋ฅผ **ํ๋์ ๋ชจ๋ธ๋ก ํฉ์ฒด**์ํค๋ ๋น๋๋ํํธ์ ๋ชจ๋ธ ์งํ ๊ธฐ์ ์
๋๋ค.
|
| 29 |
+
|
| 30 |
+
- ๊ฐ ๋ถ๋ชจ์ ์ง์์ **์์ ์์ด ๋ณด์กด**ํ ์ฑ ํฉ์นฉ๋๋ค.
|
| 31 |
+
- ๋จ์ผ ๋ถ๋ชจ๋ก๋ ๋๋ฌํ ์ ์๋ ์กฐํฉ์ ๋ง๋ญ๋๋ค โ ๊ฒฐ๊ณผ๋ฌผ์ ์ด๋ ํ ๋ถ๋ชจ๋ก๋ ํ์๋์ง ์์ต๋๋ค.
|
| 32 |
+
- **์ถ๊ฐ ์ฌ์ ํ์ต ์์ด** ์ธ๋๋ฅผ ์ด์ด๊ฐ๋๋ค.
|
| 33 |
+
|
| 34 |
+
์ด ๋ชจ๋ธ์ Chimera 4B ๊ณ๋ณด์ ๊ฒ์ฆ๋ ๊ฐ์ฒด์ ๋น๋๋ํํธ ์์ฒด **RFT(๊ฐํ ๋ฏธ์ธ์กฐ์ )** ํ์ดํ๋ผ์ธ์ ์ ์ฉํด, ํ๊ตญ์ด ์ง์ ์ถ๋ก ์ ๋์ด์ฌ๋ฆฐ ๊ฒฐ๊ณผ๋ฌผ์
๋๋ค.
|
| 35 |
+
|
| 36 |
+
---
|
| 37 |
+
|
| 38 |
+
## ์ฑ๋ฅ
|
| 39 |
+
|
| 40 |
+
**KMMLU 6๊ณผ๋ชฉ ยท 240๋ฌธํญ held-out ยท greedy ๋์ฝ๋ฉ**
|
| 41 |
+
|
| 42 |
+
| ๋ชจ๋ธ | KMMLU | vs ๊ธฐ์ค์ |
|
| 43 |
|---|---|---|
|
| 44 |
+
| ๊ธฐ์ค์ (Chimera 4B) | 42.9% | โ |
|
| 45 |
+
| **Darwin-4B-Chimera (๋ณธ ๋ชจ๋ธ)** | **48.3%** | **+5.4%p** |
|
| 46 |
+
|
| 47 |
+
๊ฐ์ ๊ฐ์ค์นยท๊ฐ์ ํ๋ค์ค์์ ์ ์ ๋จ๊ณ๋ง ๋ฌ๋ฆฌํด ์ธก์ ํ์ต๋๋ค. 6๊ณผ๋ชฉ ์ค ๋ค์์์ ์์นํ์ต๋๋ค.
|
| 48 |
+
|
| 49 |
+
---
|
| 50 |
+
|
| 51 |
+
## ์ ์งํ ๋ฒ์ (Honest scope)
|
| 52 |
+
|
| 53 |
+
์์น๋ฅผ ๋ถํ๋ฆฌ์ง ์๊ธฐ ์ํด ํ๊ณ๋ฅผ ๋จผ์ ๋ฐํ๋๋ค.
|
| 54 |
+
|
| 55 |
+
- **ํ๊ฐ ์
์ด ์์ต๋๋ค.** 240๋ฌธํญ ๊ธฐ์ค ํ์ค์ค์ฐจ ์ฝ ยฑ1.4%p์
๋๋ค. ๋ ํฐ ์
์์์ ์ฌํ์ธ์ด ํ์ํฉ๋๋ค.
|
| 56 |
+
- **KMMLU 6๊ณผ๋ชฉ ํ์ ** ์์น์ด๋ฉฐ, ์ ๋๋ฉ์ธ ์ผ๋ฐํ๋ฅผ ๋ปํ์ง ์์ต๋๋ค.
|
| 57 |
+
- **4B ๊ท๋ชจ** ๋ชจ๋ธ์
๋๋ค. ํ๋๊ทธ์ญ(Darwin-398B ๋ฑ)๊ณผ ์ง์ ๋น๊ต ๋์์ด ์๋๋๋ค.
|
| 58 |
+
- **greedy ๋์ฝ๋ฉ** ๊ธฐ์ค์
๋๋ค. ๋์ฝ๋ฉยทํ๋กฌํํธ๊ฐ ๋ฐ๋๋ฉด ๊ฒฐ๊ณผ๊ฐ ๋ฌ๋ผ์ง ์ ์์ต๋๋ค.
|
| 59 |
+
|
| 60 |
+
---
|
| 61 |
+
|
| 62 |
+
## ์ฌ์ฉ
|
| 63 |
+
|
| 64 |
+
```python
|
| 65 |
+
from transformers import AutoModelForCausalLM, AutoTokenizer
|
| 66 |
+
|
| 67 |
+
mid = "FINAL-Bench/Darwin-4B-Chimera"
|
| 68 |
+
tok = AutoTokenizer.from_pretrained(mid)
|
| 69 |
+
model = AutoModelForCausalLM.from_pretrained(mid, torch_dtype="auto", device_map="auto")
|
| 70 |
+
|
| 71 |
+
msgs = [{"role": "user", "content": "ํ๊ตญ์ ์ผ๊ถ๋ถ๋ฆฝ์ ๊ฐ๋จํ ์ค๋ช
ํด์ค."}]
|
| 72 |
+
ids = tok.apply_chat_template(msgs, add_generation_prompt=True, return_tensors="pt").to(model.device)
|
| 73 |
+
out = model.generate(ids, max_new_tokens=512, do_sample=False)
|
| 74 |
+
print(tok.decode(out[0][ids.shape[-1]:], skip_special_tokens=True))
|
| 75 |
+
```
|
| 76 |
+
|
| 77 |
+
---
|
| 78 |
+
|
| 79 |
+
## ๋ฐฉ๋ฒ๋ก ๊ณต๊ฐ ๋ฒ์
|
| 80 |
+
|
| 81 |
+
Chimera ํฉ์ฒด ๊ตฌ์ฑ๊ณผ RFT ์ ์ ์ ๋ด๋ถ ์ค๊ณ(๋ฐ์ดํฐ ์ ๋ณ ๊ธฐ์คยทํ์ต ์ค์ ยท๋ผ์ฐํ
๋ฐฉ์ ๋ฑ)๋ **๋น๋๋ํํธ ์์
๊ธฐ๋ฐ**์
๋๋ค.
|
| 82 |
+
|
| 83 |
+
> **๊ฒฐ๊ณผ๋ ๊ณต๊ฐํ๊ณ , ๋ ์ํผ๋ ๊ณต๊ฐํ์ง ์์ต๋๋ค.**
|
| 84 |
+
|
| 85 |
+
๋ชจ๋ธ ๊ฐ์ค์น์ ์ธก์ ์์น๋ ๊ณต๊ฐ๋์ด ์์ผ๋ฏ๋ก, ๋๊ตฌ๋ ์ ๊ฒฐ๊ณผ๋ฅผ ์ง์ ์ฌํยท๊ฒ์ฆํ ์ ์์ต๋๋ค.
|
| 86 |
+
|
| 87 |
+
---
|
| 88 |
+
|
| 89 |
+
## ๊ด๋ จ ๋งํฌ
|
| 90 |
+
|
| 91 |
+
- **Darwin ํจ๋ฐ๋ฆฌ** โ https://huggingface.co/collections/FINAL-Bench/darwin-family
|
| 92 |
+
- **์งํ์ ๋ณํฉ ๋ฐฉ๋ฒ๋ก ๋
ผ๋ฌธ** โ arXiv [2605.14386](https://arxiv.org/abs/2605.14386)
|
| 93 |
+
- **๋น๋๋ํํธ** โ https://vidraft.net
|
| 94 |
+
|
| 95 |
+
## ๋ผ์ด์ ์ค
|
| 96 |
|
| 97 |
+
`gemma` โ ํฉ์ฒด ๊ณ๋ณด์ Gemma ํ์ ๊ตฌ์ฑ์์๊ฐ ํฌํจ๋์ด Gemma ๋ผ์ด์ ์ค๋ฅผ ๋ฐ๋ฆ
๋๋ค. ์ฌ์ฉ ์ [Gemma Terms of Use](https://ai.google.dev/gemma/terms)๋ฅผ ํ์ธํ์ญ์์ค.
|