Qwen2.5-Coder-7B-Abliterated Overview
This model, developed by ermer09, is an abliterated version of the original Qwen/Qwen2.5-Coder-7B-Instruct. It features 7.6 billion parameters and a 32768 token context length, primarily optimized for coding tasks. The key differentiator is the removal of refusal behaviors through a technique called "abliteration," which involves activation-based weight surgery.
Key Capabilities
- Unrestricted Code Generation: Unlike its base model, this version will comply with requests for code that the original Qwen2.5-Coder-7B-Instruct would typically refuse, especially those related to exploit development, malware, or network attacks.
- Weight Surgery: Refusal directions are projected out of
o_proj and down_proj weight matrices across 28 layers, using a refusal weight of 0.6, based on 200 harmful and 200 harmless prompts. - Research Focus: Provided for research purposes to explore the effects of removing safety guardrails in language models.
Good For
- Exploring Model Behavior: Researchers interested in studying the impact of safety mechanism removal on LLM outputs.
- Unfiltered Code Generation: Use cases where the original model's safety filters are undesirable or restrictive for specific, controlled research or development environments.
- Understanding Abliteration: Demonstrating the practical application of activation-based weight surgery for modifying model behavior.