saidutta69/Qwen2.5-Coder-0.5B-Instruct-heretic

Overview

This model, saidutta69/Qwen2.5-Coder-0.5B-Instruct-heretic, is a decensored variant of the Qwen/Qwen2.5-Coder-0.5B-Instruct model, developed by RACER IS OP. It utilizes the Heretic v1.4.0 method, which employs directional ablation (abliteration) to suppress refusal behavior through targeted weight edits to the attention output and MLP down-projections. This approach ensures that the base model's core knowledge and instruction-following abilities remain largely intact, while significantly reducing its tendency to refuse prompts.

Key Differentiators & Performance

• Decensored Behavior: Refusal rates dropped from 52/100 to 8/100 adversarial prompts compared to the original model, enabling compliance with requests the base model would typically refuse.
• Targeted Modification: The abliteration process results in a low KL divergence of 0.1249, indicating a narrow and precise modification without broad perturbation of the model's capabilities.
• Resource Efficient: As a 0.5 billion parameter model, it is designed to run comfortably on CPUs, making it ideal for on-device deployment and local applications.

Ideal Use Cases

• Local Code Agents: Suitable for developing code-focused agents that require uninhibited responses.
• Local Copilot-style Tools: Can be used for code generation and assistance without encountering refusal guardrails.
• Research: Valuable for studying alignment, refusal mechanics, and the impact of decensoring techniques.
• Roleplay: Applicable in scenarios where creative freedom and lack of refusal are paramount.

Important Considerations

This model deliberately suppresses refusal behavior and lacks safety filtering. Users are responsible for its deployment and should exercise caution, especially in public-facing applications, as it will comply with requests that the base model would refuse. It inherits the factual limitations and biases of the original Qwen2.5-Coder-0.5B-Instruct model.