---
license: gemma
language:
- ko
- en
library_name: transformers
pipeline_tag: text-generation
tags:
- awaxis
- think
- gemma
- gemma-4
- reasoning
- distillation
- ko
- en
base_model:
- TeichAI/gemma-4-31B-it-Claude-Opus-Distill-v2
- google/gemma-4-31B-it
model-index:
- name: AWAXIS-Think-31B
  results:
  - task:
      type: text-generation
      name: GPQA Diamond (20Q greedy, max_new_tokens=4096)
    dataset:
      name: GPQA Diamond (subset n=20, seed=42)
      type: Idavidrein/gpqa
      config: gpqa_diamond
    metrics:
    - type: accuracy
      value: 60.0
      name: accuracy
  - task:
      type: text-generation
      name: CLIcK (Korean cultural-linguistic, n=200, alpha grid best)
    dataset:
      name: CLIcK
      type: EunsuKim/CLIcK
    metrics:
    - type: accuracy
      value: 86.0
      name: accuracy
---

# AWAXIS-Think-31B

**AWAXIS-Think-31B** is a 31B-parameter Korean/English reasoning model built via the Darwin V8 FFN-crossbreed merge engine.

## Build recipe (honest disclosure)

- **Mother (kept full)**: [TeichAI/gemma-4-31B-it-Claude-Opus-Distill-v2](https://huggingface.co/TeichAI/gemma-4-31B-it-Claude-Opus-Distill-v2) ??reasoning-distill base, retained 100% (incl. `<think>` chain-of-thought style)
- **Father (FFN donor)**: [google/gemma-4-31B-it](https://huggingface.co/google/gemma-4-31B-it) ??base Gemma-4 FFN tensors blended at **慣 = 0.1**
- **Method**: per-layer FFN blend `w = w_mother*(1-慣) + w_father*慣` on `mlp.{gate,up,down}_proj` + `pre/post_feedforward_layernorm` for all 60 language-model layers; grid search 慣??0.1, 0.2, 0.3, 0.4} on CLIcK-50 ??best 慣=0.1 (CLIcK-200 = 86.0%)
- **Architecture**: `Gemma4ForConditionalGeneration` (multimodal wrapper; text generation primary)
- **Tokenizer**: Gemma-4 (vocab 262,144)

## 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) ??*truncation artifact, included for transparency* |
| CLIcK (Korean) 200Q | greedy 慣-grid winner | 86.0% |

### Honest caveats
- GPQA 60% is from **n=20** (small sample). 16/20 still hit the 4096-token cap ??real ceiling may be higher with longer generation budget.
- Comparison to random baseline: GPQA random 25% ??+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** ??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

## Out-of-scope / limitations

- Not a final clinical/legal advisor; outputs may be confidently wrong on hard graduate-level questions (40% wrong on the GPQA-20 set).
- 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

```python
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",   # required for the Gemma4 multimodal wrapper
)
msgs = [{"role": "user", "content": "?쒓뎅?대줈 ?먯떊???뚭컻??二쇱꽭??"}]
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](https://ai.google.dev/gemma/terms).

## Acknowledgements

- TeichAI for the Opus-Distill base
- Google DeepMind for Gemma-4

---
*Built with Darwin V8 FFN-crossbreed merge engine. Measured numbers above are exact; nothing inflated.*