Multilingual Mathematical Reasoning on MGSM in-domain Rev2
90.4AccuracyAdaMame-GRPO
Evaluation Results
| Method | Links | |||
|---|---|---|---|---|
| AdaMame-GRPOBackbone=QWEN3 4B, Size=35K2026.06 | 90.4 | 88.9 | 1.9 | |
| SFTBackbone=QWEN3 4B, Size=30K2026.06 | 89.8 | 87.9 | 2.6 | |
| GRPOBackbone=QWEN3 4B, Size=35K2026.06 | 88.9 | 88.3 | 2.1 | |
| PromptBackbone=QWEN3 4B2026.06 | 88 | 14 | 14 | |
| VanillaBackbone=QWEN3 4B2026.06 | 81.3 | 13 | 14.3 | |
| AdaMame-GRPOBackbone=DISTILL-QWEN 1.5B, Size=35K2026.06 | 72.9 | 80.6 | 1.2 | |
| GRPOBackbone=DISTILL-QWEN 1.5B, Size=35K2026.06 | 72.6 | 80.1 | 1 | |
| M-Thinker Iter2Backbone=DISTILL-QWEN 1.5B, Size=50K2026.06 | 69.5 | 0.1 | 20.9 | |
| SFTBackbone=DISTILL-QWEN 1.5B, Size=30K2026.06 | 65.3 | 77.6 | 1.5 | |
| M-Thinker Iter1Backbone=DISTILL-QWEN 1.5B, Size=35K2026.06 | 62 | 0.3 | 25 | |
| PromptBackbone=DISTILL-QWEN 1.5B2026.06 | 51.6 | 57.8 | 29.4 | |
| VanillaBackbone=DISTILL-QWEN 1.5B2026.06 | 46.9 | 51.9 | 24.7 |