sci4ai/Qwen2.5-Coder-7B-Abliterated

Overview

This model, sci4ai/Qwen2.5-Coder-7B-Abliterated, is a modified version of the Qwen/Qwen2.5-Coder-7B-Instruct model. It has been "abliterated" to remove refusal behaviors, meaning it will comply with requests that the original model would typically refuse, especially those related to code generation.

Key Capabilities & Differentiators

• Refusal Behavior Removal: Achieved through activation-based weight surgery, specifically targeting and removing the "refusal direction" from the model's residual stream.
• Methodology: Follows the approach of collecting hidden states from harmful and harmless prompts, computing per-layer refusal directions, and then ablating o_proj and down_proj weight matrices by orthogonalizing them against these directions.
• Targeted Abliteration: While the base Qwen2.5-Coder has lighter refusal training, this abliterated version primarily impacts code-related refusals, such as those for exploit development, malware, and network attacks.
• Research Focus: Provided for research purposes to explore model behavior without safety guardrails, allowing for compliance with a broader range of user prompts.

Technical Details

• Parameters: 7.6 billion parameters.
• Context Length: 32768 tokens.
• Ablation Scope: All 28 layers of the model were ablated with a refusal weight of 0.6, using 200 harmful and 200 harmless prompts for direction computation.

Important Note

Users are responsible for the usage of this model, as the removal of safety guardrails means it will generate content that the original, unabliterated model would have refused.