README.md

---
language:
- ko
base_model:
- K-intelligence/Midm-2.0-Base-Instruct
---
# Midm-2.0-Base-Instruct Fine-Tuned QA Model

This model is a fine-tuned version of `Midm-2.0-Base-Instruct`, a Korean-understanding open-source LLM. It was specialized for
situation-specific Korean question answering and RAG answer explanation.

## Fine-Tuning Details

- Base model: `Midm-2.0-Base-Instruct`
- Task: Situation-specific Korean question answering
- Data type: Text
- File format: JSON
- Training data: 126,740 examples from an open Korean question-answer dataset
- Data size: 0.1 GB
- Epochs: 1

## Evaluation

| Benchmark                   | Before | After |
|-----------------------------| ---: | ---: |
| kobest_boolq (macro_f1)     | 0.865413 | 0.892661 |
| kobest_copa (macro_f1)      | 0.790682 | 0.792617 |
| kobest_hellaswag (macro_f1) | 0.506787 | 0.511349 |
| kobest_sentineg (macro_f1)  | 0.636029 | 0.640399 |
| nsmc (acc)                  | 0.626800 | 0.626267 |
| pawsx_ko (acc)              | 0.589500 | 0.596500 |

## Government R&D Project

This software is a result of the following government research and development project:

| Item | Details |
|------|---------|
| Project Name | AI 기반 중소기업 업무 자동화 및 통합 검색 솔루션 |
| Project No. | RQT-25-090153 |
| Supervisor | 정보통신산업진흥원 |
| Performer | Proten Co., Ltd. |
| R&D Period | 2026.01.01 ~ 2026.06.30 |