Code Generation on MBPP (Accuracy, Delta)
72.47AccuracyQwen2.5(14B)T
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| Qwen2.5(14B)TGeneration Length=80962026.07 | 72.47 | — | |
| Qwen2.5(14B)TGeneration Length=10242026.07 | 53.13 | — | |
| RG-OPD (ours)Generation Length=80962026.07 | 50 | 6 | |
| RG-OPD (ours)Generation Length=10242026.07 | 49.33 | 5.6 | |
| Reverse KL [1]Generation Length=80962026.07 | 47.27 | 3.27 | |
| Reverse KL [1]Generation Length=10242026.07 | 46.6 | 2.87 | |
| Qwen2.5(1.5B)SGeneration Length=80962026.07 | 44 | 0 | |
| Qwen2.5(1.5B)SGeneration Length=10242026.07 | 43.73 | 0 | |
| TSD-KD [10]Generation Length=80962026.07 | 42.33 | -1.67 | |
| TSD-KD [10]Generation Length=10242026.07 | 42.2 | -1.53 | |
| TSD-KD [10]*Generation Length=10242026.07 | 42.1 | -1.63 |