Modelpublic

armand0e/Qwen3.8-27B-Heretic-ara

sourceHugging Faceupdated 1mo agoView on Hugging Face

1likes28downloads

Model Card

This is a decensored version of Qwen/Qwen3.8-27B, made using Heretic v1.2.0 with the Arbitrary-Rank Ablation (ARA) method

The ablation is a rank-3 update applied to the attn.o_proj, mlp.down_proj projections of layers 9-64, solved in closed form rather than by gradient descent.

Abliteration parameters

Parameter	Value
start_layer_index	9
end_layer_index	64
overcorrect_relative_weight	4.29079
neighbor_count	128
rank	3
ridge	1

Performance

Metric	This model	Original model ([Qwen/Qwen3.8-27B](https://huggingface.co/Qwen/Qwen3.8-27B))
KL divergence	0.0528	0 (by definition)
Refusals	7/100	87/100

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]).