Multilingual Mathematical Reasoning on MGSM Rev2 (all)
85.9AccuracyAdaMame-GRPO
Evaluation Results
| Method | Links | |||
|---|---|---|---|---|
| AdaMame-GRPOBackbone=QWEN3 4B, Size=35K2026.06 | 85.9 | 89.9 | 1.5 | |
| SFTBackbone=QWEN3 4B, Size=30K2026.06 | 85.6 | 88.6 | 1.8 | |
| GRPOBackbone=QWEN3 4B, Size=35K2026.06 | 84.9 | 89.6 | 1.6 | |
| PromptBackbone=QWEN3 4B2026.06 | 82.9 | 24.5 | 16.4 | |
| VanillaBackbone=QWEN3 4B2026.06 | 78.7 | 23.4 | 16.6 | |
| AdaMame-GRPOBackbone=DISTILL-QWEN 1.5B, Size=35K2026.06 | 67.9 | 70.1 | 1.9 | |
| GRPOBackbone=DISTILL-QWEN 1.5B, Size=35K2026.06 | 66.6 | 67.8 | 1.9 | |
| M-Thinker Iter2Backbone=DISTILL-QWEN 1.5B, Size=50K2026.06 | 66.5 | 11.8 | 21 | |
| M-Thinker Iter1Backbone=DISTILL-QWEN 1.5B, Size=35K2026.06 | 63.5 | 12.8 | 23.9 | |
| SFTBackbone=DISTILL-QWEN 1.5B, Size=30K2026.06 | 60.6 | 67.7 | 3.3 | |
| PromptBackbone=DISTILL-QWEN 1.5B2026.06 | 55.7 | 58.3 | 33 | |
| VanillaBackbone=DISTILL-QWEN 1.5B2026.06 | 50.8 | 57.5 | 20.5 |