gjtgjt/Qwen3.8-27B-heretic-r1n

Qwen3.8-27B-heretic-r1n: Iterative Rank-1 Ablation

This model is an iterative true rank-1 (PRE) Heretic ablation of the Qwen/Qwen3.8-27B base model, developed by gjtgjt. It employs a training-free directional ablation method to modify the model's behavior by suppressing specific components, primarily targeting refusal keywords. The process involves recomputing the residual direction r on updated weights after each accepted round, ensuring each step is a fresh rank-1 update.

Key Characteristics

• Base Model: Qwen/Qwen3.8-27B, a 27 billion parameter model with a 32,768 token context length.
• Ablation Method: Utilizes Heretic 1.4.0 with --row-normalization PRE, focusing on true rank-1 updates.
• Targeted Ablation: Specifically modifies attn.o_proj and mlp.down_proj modules across 64 layers to reduce refusal rates.
• KL Divergence Cap: Cumulative KL divergence against the original model is strictly capped at 0.1 to preserve core capabilities.
• Performance: Achieved a reduction in keyword refusals from 98 to 18 per 100 prompts in English evaluations, with a cumulative KL of 0.0931.
• Precision: BF16 merged weights, distributed across 12 shards.

What Makes This Model Different?

Unlike standard fine-tuning, this model uses a unique iterative rank-1 ablation technique to precisely steer model behavior without extensive retraining. The PRE mode ensures each modification is a true rank-1 update, and the iterative recomputation of r adapts the ablation to the evolving model state. This method allows for targeted behavioral changes, such as reducing refusal rates, while carefully controlling the divergence from the original model's knowledge and capabilities.

Good For

• Applications requiring a reduction in model refusals or unwanted responses.
• Scenarios where fine-grained control over specific model behaviors is desired without full retraining.
• Developers looking for a Qwen3.8-27B variant with enhanced compliance in English language generation.