Question Answering on MedMCQA (1,000-example evaluation slice)
39.8AccuracyFull finetune
Evaluation results
| Method | Links | |
|---|---|---|
| Full finetuneBackbone=Qwen-2.5-0.5B-Instruct2026.05 | 39.8 | |
| LoRABackbone=Qwen-2.5-0.5B-Instruct2026.05 | 39 | |
| Additive sparse +SBackbone=Qwen-2.5-0.5B-Instruct, Selection rule=KL2026.05 | 37.8 | |
| Additive sparseBackbone=Qwen-2.5-0.5B-Instruct, Selection rule=KL2026.05 | 36.9 | |
| Additive sparse +SBackbone=Qwen-2.5-0.5B-Instruct, Selection rule=TF-IDF2026.05 | 36.9 | |
| Replacement sparseBackbone=Qwen-2.5-0.5B-Instruct, Selection rule=TF-IDF2026.05 | 35.5 | |
| Replacement sparseBackbone=Qwen-2.5-0.5B-Instruct, Selection rule=KL2026.05 | 35.4 | |
| Additive sparseBackbone=Qwen-2.5-0.5B-Instruct, Selection rule=TF-IDF2026.05 | 34.6 | |
| Base QwenBackbone=Qwen-2.5-0.5B-Instruct2026.05 | 34.4 |