Question Answering on Natural Questions (NQ) (EM Metric)
37.8Exact MatchDGPO
Evaluation Results
| Method | Links | |
|---|---|---|
| DGPOModel Family=Qwen 2.5, Teacher Model=Qwen 2.5 3B, Student Model=Qwen 2.5 0.5B2025.08 | 37.8 | |
| Teacher-3BModel Family=Qwen 2.5, Teacher Model=Qwen 2.5 3B2025.08 | 36.5 | |
| SFT → PPOModel Family=Qwen 2.5, Teacher Model=Qwen 2.5 3B, Student Model=Qwen 2.5 0.5B2025.08 | 33.8 | |
| DistiLLMModel Family=Qwen 2.5, Teacher Model=Qwen 2.5 3B, Student Model=Qwen 2.5 0.5B2025.08 | 33.3 | |
| SeqKDModel Family=Qwen 2.5, Teacher Model=Qwen 2.5 3B, Student Model=Qwen 2.5 0.5B2025.08 | 33.1 | |
| KDModel Family=Qwen 2.5, Teacher Model=Qwen 2.5 3B, Student Model=Qwen 2.5 0.5B2025.08 | 33.1 | |
| TAIDModel Family=Qwen 2.5, Teacher Model=Qwen 2.5 3B, Student Model=Qwen 2.5 0.5B2025.08 | 32.5 | |
| PPOModel Family=Qwen 2.5, Teacher Model=Qwen 2.5 3B, Student Model=Qwen 2.5 0.5B2025.08 | 30.6 | |
| GKDModel Family=Qwen 2.5, Teacher Model=Qwen 2.5 3B, Student Model=Qwen 2.5 0.5B2025.08 | 26.6 | |
| Student-0.5BModel Family=Qwen 2.5, Student Model=Qwen 2.5 0.5B2025.08 | 0.4 |