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

Model Overview

This model, armand0e/Qwen3.8-27B-Fable-Distill-Heretic-ara, is a 27 billion parameter large language model built upon the Qwen3.8 architecture. It is a modified version of TeichAI/Qwen3.8-27B-Fable-Distill, specifically engineered to be a "decensored" variant.

Decensoring Method

The decensoring process was performed using the Heretic v1.2.0 tool, employing the Arbitrary-Rank Ablation (ARA) method. This involved applying a rank-2 update to the attn.o_proj and mlp.down_proj projections within layers 30-64 of the original model. The ablation was solved in closed form, not via gradient descent.

Performance Highlights

• Reduced Refusals: The model demonstrates a significant reduction in refusal rates, achieving 5 refusals out of 100 harmful prompts, compared to 91 refusals for the original model. This makes it suitable for applications where the original model's content restrictions were too high.
• Low KL Divergence: Despite the modifications, the model maintains a low KL divergence of 0.1004 against the original model on harmless prompts, indicating that its general language capabilities are largely preserved.

Use Cases

This model is particularly well-suited for applications that require:

• Less restrictive content generation: Ideal for creative writing, role-playing, or research where the original model's safety filters might be overly cautious.
• Exploration of diverse outputs: Useful for developers and researchers who need to explore a broader range of model responses without frequent content refusals.