Multi-hop Question Answering on MuSiQue v1 (test)
58.1Exact Match (EM)Deepseek-V3.2
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| Deepseek-V3.2Type=API2026.02 | 58.1 | 70.97 | |
| CRAFT7BType=Trained2026.02 | 56.25 | 65.38 | |
| Gemini-2.5-FlashType=API2026.02 | 54.8 | 66.64 | |
| GPT-5-miniType=API2026.02 | 51.1 | 64.43 | |
| Qwen3-30BType=Pre-Trained2026.02 | 49.3 | 59.33 | |
| Qwen3-4BType=Pre-Trained2026.02 | 39.55 | 49.23 | |
| CRAFT3BType=Trained2026.02 | 31.8 | 42.07 | |
| Qwen2.5-7BType=Pre-Trained2026.02 | 31.55 | 41.97 | |
| Qwen2.5-7BType=Trained2026.02 | 27 | 37.82 | |
| Qwen2.5-3BType=Pre-Trained2026.02 | 23.2 | 33.31 | |
| CRAFT1.5BType=Trained2026.02 | 22.95 | 31.91 | |
| Qwen2.5-3BType=Trained2026.02 | 19.6 | 28.6 | |
| Qwen2.5-1.5BType=Pre-Trained2026.02 | 10.2 | 17.58 | |
| Qwen2.5-1.5BType=Trained2026.02 | 10.05 | 17.23 | |
| Qwen2.5-0.5BType=Pre-Trained2026.02 | 0.3 | 3.2 | |
| Qwen2.5-0.5BType=Trained2026.02 | 0.15 | 0.85 | |
| CRAFT0.5BType=Trained2026.02 | 0 | 0 |