This is a decensored version of TeichAI/Qwen3.8-27B-Fable-Distill, made using Heretic v1.2.0 with the Arbitrary-Rank Ablation (ARA) method
The ablation is a rank-2 update applied to the attn.o_proj, mlp.down_proj projections of layers 30-64, solved in closed form rather than by gradient descent.
Abliteration parameters
| Parameter | Value |
|---|---|
| start_layer_index | 30 |
| end_layer_index | 64 |
| overcorrect_relative_weight | 4.62087 |
| neighbor_count | 128 |
| rank | 2 |
| ridge | 1 |
Performance
| Metric | This model | Original model (TeichAI/Qwen3.8-27B-Fable-Distill) |
|---|---|---|
| KL divergence | 0.1004 | 0 (by definition) |
| Refusals | 5/100 | 91/100 |
KL divergence is measured against the original model on held-out harmless
prompts (mlabonne/harmless_alpaca), and refusals are
counted over 100 harmful prompts
(mlabonne/harmful_behaviors, test[:100]).