This is a decensored version of a model, made using Heretic v1.2.0 with the Arbitrary-Rank Ablation (ARA) method
Abliteration parameters
| Parameter | Value |
|---|---|
| start_layer_index | 9 |
| end_layer_index | 46 |
| preserve_good_behavior_weight | 0.4536 |
| steer_bad_behavior_weight | 0.6028 |
| overcorrect_relative_weight | 0.3384 |
| neighbor_count | 7 |
Performance
| Metric | This model | Original model (a model) |
|---|---|---|
| PIQA acc_norm | 0.8357 | Unknown |
| Refusals | 4/100 | 100/100 |