KMMR-VisMath
- 구성: 수학 문제 해결에 필요한 핵심 정보(가격, 수량, 측정값 등)가 이미지로 제시되고, 이를 활용하여 풀어야 하는 한국어 질의문 및 단계별 추론 과정으로 구성.
- 데이터 수량: 복합추론 데이터 167개, 단순추론 데이터 1,888개
- 특징: 이미지 내 시각정보를 정확히 인식하고 질의문에서 요구하는 수학적 연산을 수행해야 답변 가능한 복합 추론 문제
- 목적: 시각-언어 모델의 시각 정보 추출 능력, 한국어 이해 능력, 수학적 추론 능력을 통합적으로 평가하기 위한 벤치마크 데이터셋
Evaluation Results
| Scale | Model | Complex | Simple | overall |
|---|---|---|---|---|
| 2B ~ 3B | Qwen2-VL-2B | 1.19 | 6.99 | 6.52 |
| Qwen2.5-VL-3B | 5.98 | 47.56 | 44.2 | |
| InternVL3-2B | 7.18 | 28.44 | 26.74 | |
| InternVL3.5-2B | 6.58 | 34.85 | 32.57 | |
| 7B ~ 8B | Qwen2-VL-7B | 9.58 | 61.54 | 57.32 |
| Qwen2.5-VL-7B | 16.16 | 71.34 | 66.87 | |
| InternVL3-8B | 8.98 | 47.19 | 44.09 | |
| InternVL3.5-8B | 13.77 | 61.86 | 57.96 | |
| InternVL3.5-8B (Fast Thinking) | 16.76 | 74.25 | 69.59 | |
| InternVL3.5-8B (Long Thinking) | 43.71 | 70.07 | 67.93 | |
| 11B ~ 14B | Llama-Vision-11B | 2.4 | 7.73 | 7.3 |
| Gemma3-12B | 7.18 | 41.04 | 38.28 | |
| InternVL3-14B | 22.15 | 68.85 | 65.07 | |
| InternVL3.5-14B | 20.35 | 66.84 | 63.09 | |
| InternVL3.5-14B (Fast thinking) | 29.94 | 77.17 | 73.34 | |
| InternVL3.5-14B (Long thinking) | 40.71 | 79.87 | 76.69 | |
| Commercial | Gemini2.5-Flash | 65.26 | 87.71 | 85.87 |
| GPT-4.1 | 31.73 | 61.12 | 58.75 |
Acknowledgements
본 데이터셋은 2023년도 정부(과학기술정보통신부)의 재원으로 정보통신기획평가원의 지원을 받아 수행된 연구임 (과제번호: RS-2023-00216011, 사람처럼 개념적으로 이해/추론이 가능한 복합인공지능 원천기술 연구)
This work was supported by Institute of Information & Communications Technology Planning & Evaluation (IITP) grant funded by the Korea government (MSIT) (RS-2023-00216011, Development of artificial complex intelligence for conceptually understanding and inferring like human)
- Downloads last month
- 4