Multiple-choice Question Answering on CEval
79.86AccuracyQwen-14B
Evaluation results
| Method | Links | |
|---|---|---|
| Qwen-14Brole=Teacher model, shot=5-shot, backbone=Qwen-1.5 14B2024.07 | 79.86 | |
| + DDKrole=Student model, shot=5-shot, backbone=Qwen-1.5 1.8B, training=Knowledge Distillation (DDK)2024.07 | 65.38 | |
| + TEDrole=Student model, shot=5-shot, backbone=Qwen-1.5 1.8B, training=Knowledge Distillation (TED)2024.07 | 62.11 | |
| + MiniLLMrole=Student model, shot=5-shot, backbone=Qwen-1.5 1.8B, training=Knowledge Distillation (MiniLLM)2024.07 | 62.03 | |
| Student (1.8B)role=Student model, shot=5-shot, backbone=Qwen-1.5 1.8B2024.07 | 61.96 | |
| + KDrole=Student model, shot=5-shot, backbone=Qwen-1.5 1.8B, training=Knowledge Distillation2024.07 | 61.66 | |
| + CPTrole=Student model, shot=5-shot, backbone=Qwen-1.5 1.8B, training=Continual Pre-training2024.07 | 60.92 | |
| VisionCheckpoint selection=best2026.07 | 42.1 | |
| VisionCheckpoint selection=ep102026.07 | 42.1 | |
| IndexCheckpoint selection=ep102026.07 | 42.1 | |
| IndexCheckpoint selection=best2026.07 | 10.5 |