Commonsense Reasoning on Winogrande (Accuracy, Delta Accuracy)
88.8AccuracyLoRA
Evaluation results
| Method | Links | ||
|---|---|---|---|
| LoRABase Model=Llama-3.1-8B, Params (%)=0.26%, Rank=82026.02 | 88.8 | 2.4 | |
| Post-block steeringBase Model=Llama-3.1-8B, Params (%)=0.02%, Rank=82026.02 | 87.2 | 0.8 | |
| SFTBase Model=Llama-3.1-8B, Params (%)=100%2026.02 | 86.4 | 0 | |
| LoRABase Model=Qwen 3 4B, Params (%)=0.41%, Rank=82026.02 | 84.6 | 1.6 | |
| SFTBase Model=Qwen 3 4B, Params (%)=100%2026.02 | 83 | 0 | |
| ReFTBase Model=Llama-3.1-8B, Params (%)=0.02%, Rank=82026.02 | 82.1 | -4.5 | |
| Post-block steeringBase Model=Qwen 3 4B, Params (%)=0.04%, Rank=82026.02 | 80.6 | -2.4 | |
| ReFTBase Model=Qwen 3 4B, Params (%)=0.04%, Rank=82026.02 | 64.9 | -18.1 | |
| Post-block steeringBase Model=Llama-3.2-1B, Params (%)=0.04%, Rank=82026.02 | 60.1 | 1.9 | |
| SFTBase Model=Llama-3.2-1B, Params (%)=100%2026.02 | 58.2 | 0 | |
| LoRABase Model=Llama-3.2-1B, Params (%)=0.45%, Rank=82026.02 | 57 | -1.2 | |
| LoRABase Model=gemma-3-1b, Params (%)=0.45%, Rank=82026.02 | 51.6 | 0.2 | |
| SFTBase Model=gemma-3-1b, Params (%)=100%2026.02 | 51.4 | 0 | |
| Post-block steeringBase Model=gemma-3-1b, Params (%)=0.04%, Rank=82026.02 | 50.8 | -0.6 | |
| ReFTBase Model=Llama-3.2-1B, Params (%)=0.04%, Rank=82026.02 | 49.3 | -8.9 | |
| ReFTBase Model=gemma-3-1b, Params (%)=0.04%, Rank=82026.02 | 49.2 | -2.2 |