Procedural reasoning on Procedural-skill v1.0 (test)
95.5Performance Score (BL)pre-SFT Haiku-4-5
Evaluation results
| Method | Links | |||||
|---|---|---|---|---|---|---|
| pre-SFT Haiku-4-5Framework=API, Model Scale=Frontier, Training Stage=Pre-SFT2026.05 | 95.5 | 98.5 | 0.03 | — | — | |
| v2.0Model Scale=4B, Training Method=LoRA, Training Stage=Post-SFT2026.05 | 92 | 98.5 | 0.065 | 0.001 | 0.03 | |
| pre-SFT 4BFramework=HF, Model Scale=4B, Training Stage=Pre-SFT2026.05 | 85 | 84 | 0.01 | — | — | |
| v1.9Model Scale=2B, Training Method=LoRA, Training Stage=Post-SFT2026.05 | 81.5 | 91.5 | 0.1 | 0.0027 | 0.04 | |
| pre-SFT 2BFramework=HF, Model Scale=2B, Training Stage=Pre-SFT2026.05 | 75.5 | 81.5 | 0.06 | — | — | |
| v1Model Scale=0.8B, Training Method=LoRA, Training Stage=Post-SFT2026.05 | 71 | 70.5 | 0.005 | — | — | |
| pre-SFT 0.8BFramework=HF, Model Scale=0.8B, Training Stage=Pre-SFT2026.05 | 66.5 | 59 | 0.075 | — | — |