Multiple-choice factual recall on POLYFACT High
73.15AccuracyGRPO
Evaluation Results
| Method | Links | |
|---|---|---|
| GRPOBackbone=Qwen-2.5-7B2026.06 | 73.15 | |
| BaselineBackbone=Qwen-2.5-7B2026.06 | 66.69 | |
| GRPOBackbone=OLMo-2-1124-7B2026.06 | 64.21 | |
| CPT + GRPOBackbone=Qwen-2.5-7B2026.06 | 64.11 | |
| SFTBackbone=Qwen-2.5-7B2026.06 | 63.25 | |
| CPTBackbone=Qwen-2.5-7B2026.06 | 61.29 | |
| CPT + GRPOBackbone=OLMo-2-1124-7B2026.06 | 61.26 | |
| CPT + SFTBackbone=OLMo-2-1124-7B2026.06 | 59.2 | |
| CPT + SFTBackbone=Qwen-2.5-7B2026.06 | 58.1 | |
| BaselineBackbone=OLMo-2-1124-7B2026.06 | 57.93 | |
| CPTBackbone=OLMo-2-1124-7B2026.06 | 57.89 | |
| SFTBackbone=OLMo-2-1124-7B2026.06 | 56.33 |