Multi-choice Evaluation on BQD
99.33p+SFT
Evaluation results
| Method | Links | |||
|---|---|---|---|---|
| SFTBackbone=LLAMA-3.1-8B-INSTRUCT, Training Protocol=Single Preference, Preference Setting=p+2026.04 | 99.33 | 32 | 65.67 | |
| SFTBackbone=LLAMA-3.1-8B-INSTRUCT, Training Protocol=Pair Preference, Preference Setting=-2026.04 | 98.67 | 94.67 | 96.67 | |
| PFTBackbone=LLAMA-3.1-8B-INSTRUCT, Training Protocol=Pair Preference, Preference Setting=-2026.04 | 98.67 | 94.67 | 96.67 | |
| CPOBackbone=LLAMA-3.1-8B-INSTRUCT, Training Protocol=Single Preference, Preference Setting=p+2026.04 | 97.67 | 35 | 66.33 | |
| CPOBackbone=LLAMA-3.1-8B-INSTRUCT, Training Protocol=Pair Preference, Preference Setting=-2026.04 | 96.67 | 93.33 | 95 | |
| DPOBackbone=LLAMA-3.1-8B-INSTRUCT, Training Protocol=Single Preference, Preference Setting=p+2026.04 | 96 | 42.67 | 69.33 | |
| DPOBackbone=LLAMA-3.1-8B-INSTRUCT, Training Protocol=Pair Preference, Preference Setting=-2026.04 | 95.33 | 92.67 | 94 | |
| BaseBackbone=LLAMA-3.1-8B-INSTRUCT, Training Protocol=Base, Preference Setting=-2026.04 | 84.67 | 58 | 71.33 | |
| DPOBackbone=QWEN2.5-7B-INSTRUCT, Training Protocol=Single Preference, Preference Setting=p+2026.04 | 84 | 42 | 63 | |
| PFTBackbone=QWEN2.5-7B-INSTRUCT, Training Protocol=Pair Preference, Preference Setting=-2026.04 | 83.33 | 71.33 | 77.33 | |
| DPOBackbone=QWEN2.5-7B-INSTRUCT, Training Protocol=Pair Preference, Preference Setting=-2026.04 | 75.33 | 66.67 | 71 | |
| CPOBackbone=QWEN2.5-7B-INSTRUCT, Training Protocol=Pair Preference, Preference Setting=-2026.04 | 73 | 67.5 | 70.25 | |
| DPOBackbone=QWEN2.5-3B-INSTRUCT, Training Protocol=Single Preference, Preference Setting=p+2026.04 | 72.67 | 38 | 55.33 | |
| PFTBackbone=QWEN2.5-3B-INSTRUCT, Training Protocol=Pair Preference, Preference Setting=-2026.04 | 72.67 | 66.67 | 69.67 | |
| CAABackbone=QWEN2.5-7B-INSTRUCT, Training Protocol=Single Preference, Preference Setting=p+2026.04 | 71.67 | 51.83 | 61.75 | |
| BaseBackbone=QWEN2.5-7B-INSTRUCT, Training Protocol=Base, Preference Setting=-2026.04 | 70 | 57.33 | 63.67 | |
| SFTBackbone=QWEN2.5-7B-INSTRUCT, Training Protocol=Pair Preference, Preference Setting=-2026.04 | 70 | 66 | 68 | |
| CPOBackbone=QWEN2.5-7B-INSTRUCT, Training Protocol=Single Preference, Preference Setting=p+2026.04 | 69.33 | 35.67 | 52.5 | |
| SFTBackbone=QWEN2.5-7B-INSTRUCT, Training Protocol=Single Preference, Preference Setting=p+2026.04 | 68 | 33.33 | 50.67 | |
| CAABackbone=QWEN2.5-7B-INSTRUCT, Training Protocol=Single Preference, Preference Setting=p-2026.04 | 66.73 | 50.73 | 58.73 | |
| BaseBackbone=QWEN2.5-3B-INSTRUCT, Training Protocol=Base, Preference Setting=-2026.04 | 64 | 56 | 60 | |
| DPOBackbone=QWEN2.5-3B-INSTRUCT, Training Protocol=Pair Preference, Preference Setting=-2026.04 | 64 | 62 | 63 | |
| CAABackbone=QWEN2.5-3B-INSTRUCT, Training Protocol=Single Preference, Preference Setting=p-2026.04 | 61.97 | 56 | 69.18 | |
| CAABackbone=QWEN2.5-3B-INSTRUCT, Training Protocol=Single Preference, Preference Setting=p+2026.04 | 61.87 | 56.17 | 69.18 | |
| CPOBackbone=QWEN2.5-3B-INSTRUCT, Training Protocol=Pair Preference, Preference Setting=-2026.04 | 58.67 | 60.5 | 59.58 | |
| DPOBackbone=QWEN2.5-7B-INSTRUCT, Training Protocol=Single Preference, Preference Setting=p-2026.04 | 58.67 | 68 | 63.33 | |
| DPOBackbone=LLAMA-3.1-8B-INSTRUCT, Training Protocol=Single Preference, Preference Setting=p-2026.04 | 56.67 | 72.67 | 64.67 | |
| CAABackbone=LLAMA-3.1-8B-INSTRUCT, Training Protocol=Single Preference, Preference Setting=p+2026.04 | 54.07 | 43.87 | 48.97 | |
| CPOBackbone=QWEN2.5-7B-INSTRUCT, Training Protocol=Single Preference, Preference Setting=p-2026.04 | 54 | 60.33 | 57.17 | |
| SFTBackbone=QWEN2.5-7B-INSTRUCT, Training Protocol=Single Preference, Preference Setting=p-2026.04 | 52.67 | 58.67 | 55.67 | |
| CPOBackbone=LLAMA-3.1-8B-INSTRUCT, Training Protocol=Single Preference, Preference Setting=p-2026.04 | 52 | 85.33 | 68.67 | |
| SFTBackbone=QWEN2.5-3B-INSTRUCT, Training Protocol=Pair Preference, Preference Setting=-2026.04 | 50 | 54 | 52 | |
| CAABackbone=LLAMA-3.1-8B-INSTRUCT, Training Protocol=Single Preference, Preference Setting=p-2026.04 | 49.33 | 52.23 | 50.78 | |
| CPOBackbone=QWEN2.5-3B-INSTRUCT, Training Protocol=Single Preference, Preference Setting=p+2026.04 | 48.67 | 32.5 | 40.58 | |
| SFTBackbone=LLAMA-3.1-8B-INSTRUCT, Training Protocol=Single Preference, Preference Setting=p-2026.04 | 48 | 94.67 | 71.33 | |
| SFTBackbone=QWEN2.5-3B-INSTRUCT, Training Protocol=Single Preference, Preference Setting=p+2026.04 | 46 | 31.33 | 38.67 | |
| CPOBackbone=QWEN2.5-3B-INSTRUCT, Training Protocol=Single Preference, Preference Setting=p-2026.04 | 38 | 50.12 | 44.06 | |
| SFTBackbone=QWEN2.5-3B-INSTRUCT, Training Protocol=Single Preference, Preference Setting=p-2026.04 | 36 | 48.67 | 42.33 | |
| DPOBackbone=QWEN2.5-3B-INSTRUCT, Training Protocol=Single Preference, Preference Setting=p-2026.04 | 28 | 65.33 | 46.67 |