Temporal Question Answering on TimeQA Hard v1
0.504R-1CoT+RL pipeline
Evaluation Results
| Method | Links | |||||
|---|---|---|---|---|---|---|
| CoT+RL pipelinePrompt Type=W/ ORIGINAL CONTEXT (C), Model=Qwen2.5-1.5B, Training Method=RL2026.02 | 0.504 | 0.395 | 0.504 | 0.681 | 35.75 | |
| CoT+RL pipelinePrompt Type=W/ ORIGINAL CONTEXT (C), Model=Qwen2.5-0.5B, Training Method=RL2026.02 | 0.411 | 0.31 | 0.41 | 0.626 | 25.26 | |
| SFTPrompt Type=W/ TIME SUB-C, Model=Qwen2.5-7B2026.02 | 0.374 | 0.232 | 0.373 | 0.552 | 1.96 | |
| SFTPrompt Type=W/ ORIGINAL CONTEXT (C), Model=Qwen2.5-7B2026.02 | 0.37 | 0.18 | 0.365 | 0.54 | 0.37 | |
| SFTPrompt Type=W/ ORIGINAL CONTEXT (C), Model=Qwen2.5-1.5B2026.02 | 0.315 | 0.178 | 0.314 | 0.518 | 3.95 | |
| SFTPrompt Type=W/ TIME SUB-C, Model=Qwen2.5-1.5B2026.02 | 0.311 | 0.184 | 0.311 | 0.524 | 1.41 | |
| SFTPrompt Type=W/ KGS (WHOLE C) (FAISS), Model=Qwen2.5-7B2026.02 | 0.31 | 0.174 | 0.307 | 0.507 | 0.26 | |
| SFTPrompt Type=W/ KGS (CHUNKED C) (FAISS), Model=Qwen2.5-7B2026.02 | 0.271 | 0.164 | 0.269 | 0.484 | 0.32 | |
| SFTPrompt Type=W/ KGS (WHOLE C) (FAISS), Model=Qwen2.5-1.5B2026.02 | 0.267 | 0.18 | 0.267 | 0.462 | 0.39 | |
| LoRA (r=32)Prompt Type=W/ ORIGINAL CONTEXT (C), Model=Qwen2.5-1.5B2026.02 | 0.246 | 0.153 | 0.243 | 0.467 | 2.37 | |
| SFTPrompt Type=W/ KGS (CHUNKED C) (FAISS), Model=Qwen2.5-1.5B2026.02 | 0.225 | 0.12 | 0.225 | 0.47 | 0.13 | |
| ClassifierPrompt Type=W/ ORIGINAL CONTEXT (C), Model=Qwen2.5-7B2026.02 | 0.138 | 0.128 | 0.137 | 0.357 | 12.4 |