Reasoning on GPQADiamond SFT on Bespoke-Stratos
50.7AccuracyFull LoRA
Evaluation results
| Method | Links | |
|---|---|---|
| Full LoRABackbone=Qwen2.5-7B-Inst, k=5K2026.05 | 50.7 | |
| S^Backbone=Qwen2.5-7B-Inst, k=5K, Scoring=Gradient magnitude scoring2026.05 | 43.3 | |
| RandomBackbone=Qwen2.5-7B-Inst, k=5K2026.05 | 41.1 | |
| PretrainedBackbone=Qwen2.5-7B-Inst, k=5K2026.05 | 31.6 | |
| F^Backbone=Qwen2.5-7B-Inst, k=5K, Scoring=Diagonal Fisher scoring2026.05 | 31.1 |