Instruction Tuning on AlpacaEval 2.0 (test)
11.49Win Rate (LC)Full SFT
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| Full SFTParams (%)=100%2026.02 | 11.49 | 0.51 | |
| post-block steeringParams (%)=0.05%, mode=nonlinear, rank=162026.02 | 11.34 | 0.48 | |
| post-block steeringParams (%)=0.05%, mode=linear, rank=162026.02 | 11 | 0.53 | |
| ReFTParams (%)=0.05%, rank=16, layers=all layers2026.02 | 10.96 | 0.43 | |
| LoRAParams (%)=0.26%, rank=82026.02 | 10.52 | 0.48 | |
| LoRAParams (%)=0.52%, rank=162026.02 | 9.59 | 0.4 | |
| ReFTParams (%)=0.004%, rank=4, layers=4 layers2026.02 | 9.5 | 0.49 |