---
model-index:
  - name: EEVE-Math-10.8B
    results:
      - task:
          type: text-generation
        dataset:
          name: gsm8k-ko
          type: gsm8k
        metrics:
          - name: pass@1
            type: pass@1
            value: 0.539
            verified: false
base_model: yanolja/EEVE-Korean-10.8B-v1.0
---
# EEVE-Math-10.8B-SFT
이 모델은 [Orca-Math: Unlocking the potential of SLMs in Grade School Math](https://arxiv.org/pdf/2402.14830.pdf)및 [DARE](https://arxiv.org/abs/2311.03099)의 개념과 이를 활용한 내용을 포함하고 있습니다.

| Model | gsm8k-ko(pass@1) |
|---|---|
| Base | 0.4049 |
| SFT Epoch 1 | 0.508 |
| SFT Epoch 2(M1) | **0.539** |
| SFT -> KTO(M2) | - |
| SFT -> KTO -> KTO(final) | - |


## Specifications
SFT(M1) 단계

## Base Model
[yanolja/EEVE-Korean-10.8B-v1.0](https://huggingface.co/yanolja/EEVE-Korean-10.8B-v1.0)

## Dataset
[orca-math-word-problems-193k-korean](https://huggingface.co/datasets/kuotient/orca-math-word-problems-193k-korean)

## Evaluation
[gsm8k-ko](https://huggingface.co/datasets/kuotient/gsm8k-ko), kobest
```
git clone https://github.com/kuotient/lm-evaluation-harness.git
cd lm-evaluation-harness
pip install -e .
```
```
lm_eval --model hf \
    --model_args pretrained=yanolja/EEVE-Korean-Instruct-2.8B-v1.0 \
    --tasks gsm8k-ko \
    --device cuda:0 \
    --batch_size auto:4
```

| Model | gsm8k(pass@1) | boolq(acc) | copa(acc) | hellaswag(acc) | Overall |
|---|---|---|---|---|---|
| yanolja/EEVE-Korean-10.8B-v1.0 | 0.4049 | - | - | - | - | - |
| yanolja/EEVE-Korean-Instruct-10.8B-v1.0 | 0.4511 | **0.8668** | **0.7450** | **0.4940** | 0.6392 |
| **EEVE-Math-10.8B** | **0.5390** | 0.8027 | 0.7260 | 0.4760 | 0.6359 |
| [**EEVE-Instruct-Math-10.8B**](https://huggingface.co/kuotient/EEVE-Instruct-Math-10.8B) | 0.4951 | 0.8283 | 0.7500 | 0.4880 | **0.6403** |