AWAXIS-Think-31B

Overview

AWAXIS-Think-31B is a 31B-parameter Korean/English reasoning model created through the VIDRAFT Darwin AI Model Breeding/Evolution Platform. This model was produced using Darwin's proprietary FFN-crossbreed merge engine (V8), which emulates biological crossbreeding between AI models to create offspring with combined strengths of both parents.

AWAXIS-Think-31B์€ VIDRAFT Darwin AI ๋ชจ๋ธ ๊ต๋ฐฐ/์ง„ํ™” ํ”Œ๋žซํผ์„ ํ†ตํ•ด ์ƒ์„ฑ๋œ 31B ํŒŒ๋ผ๋ฏธํ„ฐ ํ•œ๊ตญ์–ด/์˜์–ด ์ถ”๋ก  ๋ชจ๋ธ์ž…๋‹ˆ๋‹ค.


VIDRAFT Darwin Platform

**VIDRAFT Darwin**์€ AI ๋ชจ๋ธ์˜ **๊ต๋ฐฐ(Crossbreeding)์™€ ์ง„ํ™”(Evolution)**๋ฅผ ํ†ตํ•ด ์ƒˆ๋กœ์šด ๊ณ ์„ฑ๋Šฅ ๋ชจ๋ธ์„ ์ž๋™ ์ƒ์„ฑํ•˜๋Š” ํ”Œ๋žซํผ์ž…๋‹ˆ๋‹ค. ์ƒ๋ฌผํ•™์  ์œ ์ „ ์›๋ฆฌ์—์„œ ์˜๊ฐ์„ ๋ฐ›์•„, ๋‘ ๊ฐœ ์ด์ƒ์˜ ๋ถ€๋ชจ ๋ชจ๋ธ์—์„œ ๊ฐ๊ฐ์˜ ์žฅ์ ์„ ์„ ํƒ์ ์œผ๋กœ ๊ฒฐํ•ฉํ•˜์—ฌ ์ž์‹ ๋ชจ๋ธ์„ ํƒ„์ƒ์‹œํ‚ต๋‹ˆ๋‹ค.

Darwin ๊ต๋ฐฐ/์ง„ํ™” ํ•ต์‹ฌ ๊ธฐ์ˆ 

๊ธฐ์ˆ  ์„ค๋ช…
FFN Crossbreed Engine (V8) ๋ถ€๋ชจ ๋ชจ๋ธ์˜ Feed-Forward Network(FFN) ๋ ˆ์ด์–ด๋ฅผ ์„ ํƒ์ ์œผ๋กœ ๊ต์ฐจ ๊ฒฐํ•ฉํ•˜๋Š” ํ•ต์‹ฌ ์—”์ง„. ์–ดํ…์…˜ยท์ž„๋ฒ ๋”ฉ์€ ์–ด๋จธ๋‹ˆ(Mother)์—์„œ, FFN ์‹œ๊ทธ๋„์€ ์•„๋ฒ„์ง€(Father)์—์„œ ์ถ”์ถœํ•˜์—ฌ ๋ธ”๋ Œ๋”ฉ
Smart MRI (Model Resonance Imaging) ๋‘ ๋ชจ๋ธ ๊ฐ„ ๋ ˆ์ด์–ด๋ณ„ ์œ ์‚ฌ๋„ยทํ˜ธํ™˜์„ฑ์„ ๋ถ„์„ํ•˜์—ฌ ์ตœ์  ๊ต๋ฐฐ ๋น„์œจ(alpha)์„ ์ž๋™ ํƒ์ƒ‰ํ•˜๋Š” ๊ธฐ์ˆ 
Alpha Grid Search ๊ต๋ฐฐ ๋น„์œจ alpha๋ฅผ ์ฒด๊ณ„์ ์œผ๋กœ ํƒ์ƒ‰(0.1~0.4)ํ•˜์—ฌ ๋ฒค์น˜๋งˆํฌ ์„ฑ๋Šฅ์ด ์ตœ๋Œ€ํ™”๋˜๋Š” ์ตœ์ ์ ์„ ๋ฐœ๊ฒฌ
Multi-Generation Breeding 1์„ธ๋Œ€ ๊ต๋ฐฐ ๊ฒฐ๊ณผ๋ฌผ์„ ๋‹ค์‹œ ๋ถ€๋ชจ๋กœ ์‚ผ์•„ 2์„ธ๋Œ€, 3์„ธ๋Œ€ ๊ต๋ฐฐ๋ฅผ ์ˆ˜ํ–‰ํ•˜๋Š” ๋‹ค์„ธ๋Œ€ ์ง„ํ™”

Darwin ๊ต๋ฐฐ ํ”„๋กœ์„ธ์Šค (์ด ๋ชจ๋ธ์˜ ์ƒ์„ฑ ๊ณผ์ •)

[Step 1] ๋ถ€๋ชจ ์„ ์ • (Parent Selection)
   - Mother: TeichAI/gemma-4-31B-it-Claude-Opus-Distill-v2 (์ถ”๋ก  ๋Šฅ๋ ฅ ๊ธฐ๋ฐ˜)
   - Father: google/gemma-4-31B-it (Gemma-4 ์›๋ณธ FFN ๊ธฐ์—ฌ)

[Step 2] Smart MRI ํ˜ธํ™˜์„ฑ ๋ถ„์„
   - ๋‘ ๋ชจ๋ธ์˜ 60๊ฐœ ๋ ˆ์ด์–ด๋ณ„ FFN ํ…์„œ ์œ ์‚ฌ๋„ ์Šค์บ”
   - ์•„ํ‚คํ…์ฒ˜ ํ˜ธํ™˜์„ฑ ํ™•์ธ (๋™์ผ Gemma-4 family = 100% ํ˜ธํ™˜)

[Step 3] FFN Crossbreed (๊ต๋ฐฐ ์‹คํ–‰)
   - ์–ด๋จธ๋‹ˆ์˜ ์–ดํ…์…˜, ์ž„๋ฒ ๋”ฉ, ๋ผ์šฐํŒ… = 100% ๋ณด์กด
   - ์•„๋ฒ„์ง€์˜ FFN (gate_proj, up_proj, down_proj) = alpha ๋น„์œจ๋กœ ๋ธ”๋ Œ๋”ฉ
   - ์ˆ˜์‹: w_child = w_mother * (1 - alpha) + w_father * alpha

[Step 4] Alpha Grid Search (์ตœ์  ๊ต๋ฐฐ ๋น„์œจ ํƒ์ƒ‰)
   - alpha = {0.1, 0.2, 0.3, 0.4} 4์ข… ์ƒ์„ฑ
   - CLIcK-50 ๋ฒค์น˜๋งˆํฌ๋กœ ๊ฐ alpha ํ‰๊ฐ€
   - ์ตœ์ : alpha = 0.1 (CLIcK-200 = 86.0%)

[Step 5] ๊ฒ€์ฆ ๋ฐ ์ถœ์‹œ
   - GPQA Diamond, CLIcK ๋“ฑ ๋ฒค์น˜๋งˆํฌ ๊ฒ€์ฆ
   - HuggingFace ๋ชจ๋ธ ํ—ˆ๋ธŒ ๊ณต๊ฐœ

์™œ Darwin ๊ต๋ฐฐ์ธ๊ฐ€?

๊ธฐ์กด ๋ชจ๋ธ ํ•ฉ์„ฑ ๋ฐฉ์‹(๋‹จ์ˆœ ๊ฐ€์ค‘์น˜ ํ‰๊ท , SLERP, TIES ๋“ฑ)๊ณผ ๋‹ฌ๋ฆฌ, Darwin ๊ต๋ฐฐ๋Š”:

  1. ์ƒ๋ฌผํ•™์  ์œ ์ „ ๋ชจ๋ฐฉ: ์–ด๋จธ๋‹ˆ/์•„๋ฒ„์ง€ ์—ญํ• ์„ ๋ช…ํ™•ํžˆ ๋ถ„๋ฆฌํ•˜์—ฌ ๊ฐ ๋ถ€๋ชจ์˜ ํ•ต์‹ฌ ๋Šฅ๋ ฅ๋งŒ ์„ ํƒ์ ์œผ๋กœ ์ƒ์†
  2. FFN ์„ ํƒ์  ์ฃผ์ž…: ์–ดํ…์…˜(๋ฌธ๋งฅ ์ดํ•ด)์€ ์–ด๋จธ๋‹ˆ์—์„œ 100% ๋ณด์กดํ•˜๊ณ , FFN(์ง€์‹ยท์ถ”๋ก  ํŒจํ„ด)๋งŒ ์•„๋ฒ„์ง€์—์„œ ๊ต์ฐจ โ†’ ๋Šฅ๋ ฅ ์ถฉ๋Œ ์ตœ์†Œํ™”
  3. ๋ฒค์น˜๋งˆํฌ ๊ธฐ๋ฐ˜ ์ž์—ฐ์„ ํƒ: alpha grid search๋กœ ์—ฌ๋Ÿฌ ์ž์‹ ํ›„๋ณด๋ฅผ ์ƒ์„ฑํ•œ ๋’ค, ์‹ค์ธก ๋ฒค์น˜๋งˆํฌ๋กœ ์ตœ์  ๊ฐœ์ฒด๋ฅผ ์„ ํƒ (= ์ž์—ฐ์„ ํƒ ์‹œ๋ฎฌ๋ ˆ์ด์…˜)
  4. ๋‹ค์„ธ๋Œ€ ์ง„ํ™” ๊ฐ€๋Šฅ: ์ด ๋ชจ๋ธ(AWAXIS-Think-31B)์ด ๋‹ค์‹œ AWAXIS-KR-31B์˜ ์•„๋ฒ„์ง€๊ฐ€ ๋˜์–ด 2์„ธ๋Œ€ ๊ต๋ฐฐ ์ˆ˜ํ–‰ โ†’ ๋Šฅ๋ ฅ ๋ˆ„์  ์ง„ํ™”

Build Recipe (Honest Disclosure)

  • Mother (kept full): TeichAI/gemma-4-31B-it-Claude-Opus-Distill-v2 โ€” reasoning-distill base, retained 100% (incl. <think> chain-of-thought style)
  • Father (FFN donor): google/gemma-4-31B-it โ€” base Gemma-4 FFN tensors blended at alpha = 0.1
  • Method: Darwin V8 FFN-crossbreed โ€” per-layer FFN blend w = w_mother*(1-alpha) + w_father*alpha on mlp.{gate,up,down}_proj + pre/post_feedforward_layernorm for all 60 language-model layers; grid search alpha in {0.1, 0.2, 0.3, 0.4} on CLIcK-50 โ€” best alpha=0.1 (CLIcK-200 = 86.0%)
  • Platform: VIDRAFT Darwin AI Model Breeding/Evolution Platform (VIDraft on HuggingFace)
  • Architecture: Gemma4ForConditionalGeneration (multimodal wrapper; text generation primary)
  • Tokenizer: Gemma-4 (vocab 262,144)

Model Lineage (Genealogy)

AWAXIS-Think-31B  (this model -- Darwin V8 FFN-crossbreed)
|
+-- Mother (kept full, 100%)
|   TeichAI/gemma-4-31B-it-Claude-Opus-Distill-v2
|   -- Claude Opus reasoning distill base
|
+-- Father (FFN donor, alpha=0.1)
    google/gemma-4-31B-it
    -- Gemma-4 base FFN tensors

Common ancestor: Google Gemma-4 architecture.


Measured Benchmarks

Benchmark Setting Result
GPQA Diamond 20Q (seed 42) greedy, max_new_tokens=4096, 2-way DP 12/20 = 60.0% (16/20 still hit token cap, 0 null)
GPQA Diamond 20Q (seed 42) greedy, max_new_tokens=2048 9/20 = 45.0% (16/20 truncated, 2 null) โ€” truncation artifact, included for transparency
CLIcK (Korean) 200Q greedy alpha-grid winner 86.0%

Honest Caveats

  • GPQA 60% is from n=20 (small sample). 16/20 still hit the 4096-token cap โ€” real ceiling may be higher with longer generation budget.
  • Comparison to random baseline: GPQA random 25% โ€” +35pp clear learning signal.
  • The full GPQA Diamond (198Q) and other broad suites have not yet been measured for this exact merged artifact.
  • The model retains the Mother's <think>...</think> reasoning template โ€” strip via post-processing if undesired.

Intended Use

  • Korean/English step-by-step reasoning, instruction following, knowledge QA
  • The Think suffix reflects the inherited Opus-distilled chain-of-thought behavior
  • 2nd-generation breeding parent: This model served as the Father for AWAXIS-KR-31B, demonstrating Darwin's multi-generation evolution capability

Out-of-Scope / Limitations

  • Not a final clinical/legal advisor; outputs may be confidently wrong on hard graduate-level questions
  • Inherits Gemma-4 base limitations (multimodal wrapper retained; image inputs not the primary use-case here)
  • Subject to Gemma Terms of Use; see parent model cards for derivative-use clauses

Inference

from transformers import AutoTokenizer, AutoModelForCausalLM
import torch

tok = AutoTokenizer.from_pretrained("Anserwise/AWAXIS-Think-31B", trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(
    "Anserwise/AWAXIS-Think-31B",
    dtype=torch.bfloat16,
    device_map="auto",
    trust_remote_code=True,
    attn_implementation="eager",
)

msgs = [{"role": "user", "content": "์–‘์ž์—ญํ•™์˜ ๋ถˆํ™•์ •์„ฑ ์›๋ฆฌ๋ฅผ ์ž์„ธํžˆ ์„ค๋ช…ํ•ด ์ฃผ์„ธ์š”."}]
text = tok.apply_chat_template(msgs, tokenize=False, add_generation_prompt=True)
inp = tok(text, return_tensors="pt").to(model.device)
out = model.generate(**inp, max_new_tokens=2048, do_sample=False)
print(tok.decode(out[0][inp["input_ids"].shape[-1]:], skip_special_tokens=True))

License

Gemma Terms of Use (inherited from base). Use of this model is bound by Google Gemma Terms.

Acknowledgements

  • VIDRAFT โ€” Darwin AI Model Breeding/Evolution Platform
  • TeichAI for the Opus-Distill base
  • Google DeepMind for Gemma-4

Built with the VIDRAFT Darwin AI Model Breeding/Evolution Platform โ€” FFN-crossbreed V8 engine. This model was generated through Darwin's automated crossbreeding process, which selectively combines the strengths of parent models using biologically-inspired genetic algorithms. Measured numbers above are exact; nothing inflated.

Downloads last month
18
Safetensors
Model size
31B params
Tensor type
BF16
ยท
Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐Ÿ™‹ Ask for provider support

Model tree for Anserwise/AWAXIS-Think-31B

Finetuned
(1)
this model
Finetunes
2 models
Quantizations
2 models

Evaluation results