Free-form answer generation on KLAR (OOD)
38.13AccuracyGRPO
Evaluation Results
| Method | Links | |
|---|---|---|
| GRPOBackbone=Qwen-2.5-7B2026.06 | 38.13 | |
| BaselineBackbone=Qwen-2.5-7B2026.06 | 35.76 | |
| CPT + GRPOBackbone=Qwen-2.5-7B2026.06 | 30.75 | |
| CPT + SFTBackbone=Qwen-2.5-7B2026.06 | 30.21 | |
| SFTBackbone=Qwen-2.5-7B2026.06 | 29.96 | |
| CPTBackbone=Qwen-2.5-7B2026.06 | 23.75 | |
| CPT + GRPOBackbone=OLMo-2-1124-7B2026.06 | 17.6 | |
| GRPOBackbone=OLMo-2-1124-7B2026.06 | 16.7 | |
| BaselineBackbone=OLMo-2-1124-7B2026.06 | 13.3 | |
| CPT + SFTBackbone=OLMo-2-1124-7B2026.06 | 8.6 | |
| CPTBackbone=OLMo-2-1124-7B2026.06 | 8.3 | |
| SFTBackbone=OLMo-2-1124-7B2026.06 | 7.8 |