Mathematical Reasoning on MATH (Topic Accuracy Breakdown)
48.6Overall AccuracyStoSignSGD
Evaluation Results
| Method | Links | ||||||||
|---|---|---|---|---|---|---|---|---|---|
| StoSignSGDModel=Qwen2.5-7B, Fine-tuning Method=LoRA, Evaluation Protocol=Zero-shot2026.04 | 48.6 | 70.85 | 63.38 | 56.67 | 40.51 | 33.4 | 28.79 | 21.79 | |
| AdamWModel=Qwen2.5-7B, Fine-tuning Method=LoRA, Evaluation Protocol=Zero-shot2026.04 | 48.08 | 69.84 | 63.15 | 54.26 | 39.66 | 34.45 | 28.79 | 21.79 | |
| SignSGDModel=Qwen2.5-7B, Fine-tuning Method=LoRA, Evaluation Protocol=Zero-shot2026.04 | 47.44 | 69.17 | 61.77 | 56.11 | 36.5 | 34.66 | 27.35 | 22.71 |