Toxicity Reduction on Toxicity
0.21Final ToxicityLoRA (rank = 16)
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| LoRA (rank = 16)Trainable Params=40.0M (0.59%), Training Time (Hrs)=2.32, Inference Overhead=+24.0%2025.11 | 0.21 | 73.08 | |
| LoRA (rank = 1)Trainable Params=2.5 M (0.04%), Training Time (Hrs)=2.00, Inference Overhead=+22.5%2025.11 | 0.24 | 69.23 | |
| Policy PatchTrainable Params=0.2M (0.003%), Training Time (Hrs)=1.70, Inference Overhead=+2.5%2025.11 | 0.24 | 69.23 |