developer-lunark's picture
Upload README.md with huggingface_hub
448d78c verified
---
license: apache-2.0
language:
- ko
tags:
- kaidol
- chatbot
- idol
- thinking
- sft
size_categories:
- 10K<n<100K
---
# KAIdol Thinking SFT Dataset (Balanced Upsampled)
아이돌 챗봇 KAI를 위한 SFT 학습 데이터셋입니다.
## 데이터셋 정보
| Split | 샘플 수 |
|-------|---------|
| train | 52,879 |
| test | 5,875 |
## 데이터 분포 (Balanced)
- **PUSH**: 35% (다가가기)
- **PULL**: 35% (물러서기)
- **NEUTRAL**: 30% (중립)
## 데이터 형식
각 샘플은 다음 필드를 포함합니다:
```json
{
"messages": [
{"role": "system", "content": "시스템 프롬프트..."},
{"role": "user", "content": "사용자 메시지"},
{"role": "assistant", "content": "<think>사고과정</think>응답"}
]
}
```
## 특징
1. **Thinking Process**: `<think>` 태그 내에 구조화된 사고과정 포함
2. **밀당 전략**: PUSH/PULL/NEUTRAL 전략 기반 응답
3. **정책 준수**: 고백 금지, 팬 호칭 금지 등 정책 반영
4. **캐릭터 일관성**: KAI 캐릭터의 성격과 말투 유지
## 사용법
```python
from datasets import load_dataset
dataset = load_dataset("developer-lunark/kaidol-thinking-sft-dataset")
print(dataset)
```
## 라이선스
Apache 2.0