Puerh0x1/Qwen3.5-9B-abliterated
Puerh0x1/Qwen3.5-9B-abliterated is a 9 billion parameter language model based on Qwen/Qwen3.5-9B, featuring a 32768 token context length. This model is a weight-space edit, not a fine-tune, specifically engineered to ablate refusal directions. It aims to reduce model refusals through a targeted modification process rather than traditional instruction tuning, making it suitable for applications requiring less restrictive content generation.
Loading preview...
Qwen3.5-9B-abliterated Overview
Puerh0x1/Qwen3.5-9B-abliterated is a 9 billion parameter language model derived from the Qwen/Qwen3.5-9B base model. Unlike traditional fine-tuning, this model represents a weight-space edit, specifically designed to ablate refusal directions. The primary goal of this modification is to reduce the model's tendency to refuse certain prompts, without claiming complete removal of refusal behaviors.
Key Characteristics
- Weight-Space Edit: This model is a direct modification of the base model's weights, not a fine-tuned version.
- Refusal-Direction Ablation: Engineered to reduce refusals, as indicated by validation measurements showing 0/32 keyword hits on validation refusals and a low KL divergence on harmless prompts.
- Base Model: Built upon Qwen/Qwen3.5-9B, maintaining its 32768 token context length.
- Methodology: Utilizes the Heretic tool with BF16 precision and contrastive prompts from Arditi et al. 2024 for targeted weight modification.
- Vision Weights Unchanged: The vision-related weights from the base model were not altered during this process.
Use Cases
- Content Generation: Potentially useful for applications where a reduced tendency for model refusals is desired.
- Research into Model Safety: Can serve as a case study for understanding and mitigating refusal behaviors through weight-space editing.
- Experimental Deployments: Suitable for developers exploring alternative methods of controlling model outputs beyond standard instruction tuning.