README.md

---
license: cc-by-nc-4.0
datasets:
- kyujinpy/KOR-OpenOrca-Platypus-v3
- beomi/KoAlpaca-v1.1a
- maywell/ko_wikidata_QA
language:
- ko
base_model: beomi/Yi-Ko-6B
---

# Yi-Ko-6B-Instruct-v1.0

## Model Details

### Base Model
[beomi/Yi-Ko-6B](https://huggingface.co./beomi/Yi-Ko-6B)

### Training Dataset
1. [kyujinpy/KOR-OpenOrca-Platypus-v3](https://huggingface.co./datasets/kyujinpy/KOR-OpenOrca-Platypus-v3) 🙇
2. [beomi/KoAlpaca-v1.1a](https://huggingface.co./datasets/beomi/KoAlpaca-v1.1a) 🙇
3. [maywell/ko_wikidata_QA](https://huggingface.co./datasets/maywell/ko_wikidata_QA) 🙇
4. AIHub MRC 데이터 선별 후 Instruction Format 맞게 변경 후 사용

## Benchmark Results

### AI-Harness Evaluation  
https://github.com/Beomi/ko-lm-evaluation-harness 
  
| Model | kobest_boolq | kobest_copa | kobest_hellaswag | kobest_sentineg | korunsmile | pawsx_ko |  
| --- | --- | --- | --- | --- | --- | --- |
|  | *Zero-shot* ||||||
| Yi-Ko-6B-Instruct-v1.0 | 0.6619 | 0.7794 | 0.4858 | 0.4589 | 0.3520 | 0.5545 |
| Yi-Ko-6B | 0.7070 | 0.7696 | 0.5009 | 0.4044 | 0.3828 | 0.5145 |

## Instruction Format
```python
### User:
{instruction}

### Assistant:
{response}
```

## Loading the Model
```python
import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

tokenizer = AutoTokenizer.from_pretrained("wkshin89/Yi-Ko-6B-Instruct-v1.0")
model = AutoModelForCausalLM.from_pretrained(
    "wkshin89/Yi-Ko-6B-Instruct-v1.0",
    device_map="auto",
    torch_dtype=torch.bfloat16,
)
```