Math Word Problem Solving on GSM+ v1 (test)
65.7AccuracyFull-FT
Evaluation Results
| Method | Links | |
|---|---|---|
| Full-FTBackbone=Qwen3-8B-Base, KV Sharing=Not Supported2026.02 | 65.7 | |
| PrefillShareBackbone=Qwen3-8B-Base, KV Sharing=Supported2026.02 | 64.5 | |
| Full-FTBackbone=LLaMA3.1-8B, KV Sharing=Not Supported2026.02 | 49.8 | |
| PrefillShareBackbone=LLaMA3.1-8B, KV Sharing=Supported2026.02 | 49.3 | |
| LLaMA3.1-8BKV Sharing=Inherent2026.02 | 18 | |
| Qwen3-8B-BaseKV Sharing=Inherent2026.02 | 12.5 |