Mathematical Reasoning on GSM8K (Speedup ratio)
2.414Speedup RatioMTP-D ensemble
Evaluation results
| Method | Links | |
|---|---|---|
| MTP-D ensembleLoop strategy=4 to 8, Training protocol=Continued pre-training, Training data size=70B tokens2026.03 | 2.414 | |
| MTP-DLoop strategy=4 to 8, Training protocol=Continued pre-training, Training data size=350B tokens2026.03 | 2.393 | |
| MTP-DLoop strategy=4 to 8, Training protocol=Continued pre-training, Training data size=70B tokens2026.03 | 2.365 | |
| MTPLoop strategy=4 to 8, Training protocol=Continued pre-training, Training data size=70B tokens2026.03 | 2.23 | |
| MTP-DLoop strategy=1 to 8, Training protocol=Continued pre-training, Training data size=70B tokens2026.03 | 2.197 | |
| MTP-DLoop strategy=4, Training protocol=Continued pre-training, Training data size=70B tokens2026.03 | 2.107 | |
| MTP-DLoop strategy=4 to 8, Training protocol=Training Free, Training data size=70B tokens2026.03 | 2.058 | |
| MTPLoop strategy=1 to 8, Training protocol=Continued pre-training, Training data size=70B tokens2026.03 | 2.034 | |
| MTP-DLoop strategy=4 to 16, Training protocol=Continued pre-training, Training data size=70B tokens2026.03 | 2.012 | |
| MTP-DLoop strategy=4 to 8 to 16, Training protocol=Continued pre-training, Training data size=70B tokens2026.03 | 1.958 | |
| MTP-DLoop strategy=4 to 16, Training protocol=Training Free, Training data size=70B tokens2026.03 | 1.713 | |
| MTP-DLoop strategy=1 to 16, Training protocol=Continued pre-training, Training data size=70B tokens2026.03 | 1.698 | |
| MTP-DLoop strategy=1 to 8, Training protocol=Training Free, Training data size=70B tokens2026.03 | 1.451 | |
| MTP-DLoop strategy=1 to 16, Training protocol=Training Free, Training data size=70B tokens2026.03 | 1.156 | |
| MTP-DLoop strategy=1, Training protocol=Continued pre-training, Training data size=70B tokens2026.03 | 1.142 |