Math Reasoning on NumGLUE, MMLU-Math, GSM8K, SVAMP, SimulEq, DeepMind, AQuA, SAT
57.68NumGLUE AccuracyID
Evaluation results
| Method | Links | |||||||||
|---|---|---|---|---|---|---|---|---|---|---|
| IDBackbone=Qwen2.5-7B, Fine-tuning Dataset=MetaMathQA, Fine-tuning Steps=2048, Full Batch Size=128, Selection Budget (κ)=16, Subset Ratio=12.5%2026.06 | 57.68 | 64.68 | 80.52 | 84.6 | 41.44 | 37.9 | 60.24 | 78.18 | 63.15 | |
| PartitionSelBackbone=Qwen2.5-7B, Fine-tuning Dataset=MetaMathQA, Fine-tuning Steps=2048, Full Batch Size=128, Selection Budget (κ)=16, Subset Ratio=12.5%2026.06 | 56.43 | 64.58 | 83.47 | 86.4 | 38.72 | 38.1 | 63.78 | 83.64 | 64.39 |