DARWIN-Family
Collection
๋น๋๋ํํธ โข 57 items โข Updated โข 29
Darwin ร Mistral ์งํ์ ๋จธ์ง 1์ธ๋ (G1) โ Devstral-2-123B(์ฝ๋ฉ) ร Mistral-Medium-3.5-128B(์ถ๋ก ยท๋ฒ์ฉ) ์ Darwin V7+ MRI-only Linear merge ๊ฒฐ๊ณผ๋ฌผ.
VIDRAFT R&D ยท 2026-05-16 ์์ ์ง์์ v1.0 (๊นํ๋ด CAIO) ์ต์ A ์ฑํ ๊ฒฐ๊ณผ
| Benchmark | Score | Note |
|---|---|---|
| GPQA Diamond 20 (seed 42) | 45.0% (9/20) | Mother ์์ค ํ๋ณต โ Devstral 50% / MM3.5 45% ์ฌ์ด |
| HumanEval 20 pass@1 (seed 42) | ๐ฅ 90.0% (18/20) | Devstral ์ฝ๋ฉ ๋ฅ๋ ฅ ์๋ฒฝ ๋ณด์กด |
| Sanity (์ํยท์ฝ๋ยทํ๊ตญ์ด) | โ 3/3 | 17ร23=391, fibonacci code, ํ๊ตญ์ด ๋ฒ์ญ ์ ์ |
mistralai/Devstral-2-123B-Instruct-2512 โ Ministral3ForCausalLM, hidden 12288, layers 88, vocab 131072, FP8 + BF16mistralai/Mistral-Medium-3.5-128B โ Mistral3ForConditionalGeneration text branch (vision ์ ์ธ), ๋์ผ spec(1 - r_mother) * Father + r_mother * Mother per tensormodel.language_model.X โ model.X (Ministral3 namespace ํต์ผ)| Layer ๊ตฌ๊ฐ | r_mother | ์๋ |
|---|---|---|
| 0~29 (low) | 0.25 | Father(Devstral) ์ฐ์ธ โ ํ ํฐ ํํยท์ฝ๋ ๊ตฌ๋ฌธ ๋ณด์กด |
| 30~59 (mid) | 0.70 | Mother ์ฐ์ธ โ ์ถ๋ก ์ฒด์ธ ๊ฐํ |
| 60~87 (high) | 0.50 | ๊ท ๋ฑ โ ์ถ๋ ฅ ํ์งยท์ธ์คํธ๋ญ์ |
| Category | r_mother | ๋น๊ณ |
|---|---|---|
| embedding | 0.00 | Father 100% โ ํ ํฐ ๋งคํ ๋ณด์กด (v1 1.00 ์ random ์์ ํ์ธ) |
| lm_head | 0.00 | Father 100% โ ์ถ๋ ฅ ๋งคํ ๋ณด์กด |
| attention | 0.50 | ๊ท ๋ฑ โ Devstral reasoning ํ๋ณต |
| norm | 0.50 | ๊ท ๋ฑ โ LayerNorm ํต๊ณ |
| router/shared_expert/routed_expert | n/a | Mistral dense๋ผ ํด๋น ์์ |
| ๋ชจ๋ธ | ์ ๋ต/์ด | ์ ํ๋ |
|---|---|---|
| Devstral-2-123B (Father) | 10/20 | 50.0% |
| Mistral-Medium-3.5-128B (Mother) | 9/20 | 45.0% |
| Darwin-Mistral-G1 v1 (failed) | 5/20 | 25.0% (random) |
| Darwin-Mistral-G1 v2 (this) | 9/20 | 45.0% โ |
V1 ์คํจ ์์ธ (GPQA 25% random level):
V2 fix:
from transformers import AutoTokenizer, AutoModelForCausalLM
import torch
# ์์กด์ฑ: pip install --user kernels (transformers 5.5.x finegrained-fp8 kernel)
model = AutoModelForCausalLM.from_pretrained(
"FINAL-Bench/Darwin-Mistral-G1",
torch_dtype=torch.bfloat16,
device_map="auto",
trust_remote_code=True,
)
tok = AutoTokenizer.from_pretrained("FINAL-Bench/Darwin-Mistral-G1")
messages = [{"role": "user", "content": "Write a Python function to compute Fibonacci."}]
text = tok.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inp = tok(text, return_tensors="pt").to(model.device)
out = model.generate(**inp, max_new_tokens=256, do_sample=False)
print(tok.decode(out[0][inp.input_ids.shape[1]:], skip_special_tokens=True))
ํ๊ฒฝ ์๊ตฌ์ฌํญ:
transformers >= 5.5.0kernels >= 0.14.0 (finegrained-fp8 kernel ์๋ ๋ค์ด๋ก๋)torch >= 2.10 (FP8 E4M3 ์ง์)Apache-2.0 (Devstral-2-123B-Instruct-2512์ Mistral-Medium-3.5-128B ๋ชจ๋ Apache 2.0). ๋จธ์ง ์์ง์ VIDRAFT Darwin V7+ MRI-only (proprietary).
@misc{darwin-mistral-g1-2026,
title = {Darwin-Mistral-G1: Linear MRI Merge of Devstral-2-123B and Mistral-Medium-3.5-128B},
author = {VIDRAFT R&D},
year = {2026},
url = {https://huggingface.co/FINAL-Bench/Darwin-Mistral-G1}
}
VIDRAFT Inc. | ๊นํ๋ด CAIO | CONFIDENTIAL (private repo until publication)