Multilingual Mathematical Reasoning on MGSM out-of-domain Rev2
83.6AccuracySFT
Evaluation Results
| Method | Links | |||
|---|---|---|---|---|
| SFTBackbone=QWEN3 4B, Size=30K2026.06 | 83.6 | 88.9 | 1.8 | |
| AdaMame-GRPOBackbone=QWEN3 4B, Size=35K2026.06 | 83.6 | 90.5 | 1.4 | |
| GRPOBackbone=QWEN3 4B, Size=35K2026.06 | 82.9 | 90.3 | 1.6 | |
| PromptBackbone=QWEN3 4B2026.06 | 80.4 | 36 | 17.6 | |
| VanillaBackbone=QWEN3 4B2026.06 | 77.4 | 34.5 | 17.8 | |
| AdaMame-GRPOBackbone=DISTILL-QWEN 1.5B, Size=35K2026.06 | 65.5 | 64.8 | 2.3 | |
| M-Thinker Iter2Backbone=DISTILL-QWEN 1.5B, Size=50K2026.06 | 65 | 17.7 | 21.1 | |
| M-Thinker Iter1Backbone=DISTILL-QWEN 1.5B, Size=35K2026.06 | 64.2 | 19 | 23.3 | |
| GRPOBackbone=DISTILL-QWEN 1.5B, Size=35K2026.06 | 63.6 | 61.6 | 2.6 | |
| SFTBackbone=DISTILL-QWEN 1.5B, Size=30K2026.06 | 58.3 | 62.8 | 4.2 | |
| PromptBackbone=DISTILL-QWEN 1.5B2026.06 | 57.8 | 58.6 | 34.7 | |
| VanillaBackbone=DISTILL-QWEN 1.5B2026.06 | 52.8 | 60.3 | 18.4 |