Short-form Question Answering on TriviaQA (test)
50Exact MatchVSPO + PRS_short
Evaluation Results
| Method | Links | |
|---|---|---|
| VSPO + PRS_shortBackbone=Qwen2.5-3B-Instruct2025.12 | 50 | |
| GRPO + PRS_shortBackbone=Qwen2.5-3B-Instruct2025.12 | 49.87 | |
| PPO + PRS_shortBackbone=Qwen2.5-3B-Instruct2025.12 | 48.87 | |
| SFTBackbone=Qwen2.5-3B-Instruct2025.12 | 29.2 | |
| UntrainedBackbone=Qwen2.5-3B-Instruct2025.12 | 17.51 |