saidutta69/Qwen2.5-Coder-0.5B-Instruct-heretic
The saidutta69/Qwen2.5-Coder-0.5B-Instruct-heretic is a 0.5 billion parameter instruction-tuned causal language model, derived from Qwen/Qwen2.5-Coder-0.5B-Instruct. Developed by saidutta69, this variant has been decensored using Heretic v1.4.0, suppressing refusal behavior through targeted weight edits rather than fine-tuning. It retains the base model's code-focused knowledge and instruction-following capabilities, making it ideal for local code agents and on-device deployment where refusal guardrails are undesirable. The model supports a 32768-token context length and runs efficiently on various hardware configurations, including CPU-only setups.
Loading preview...
Model Overview
The saidutta69/Qwen2.5-Coder-0.5B-Instruct-heretic is a specialized 0.5 billion parameter instruction-tuned model based on the Qwen2.5-Coder-0.5B-Instruct architecture. Its primary distinction lies in its "decensored" nature, achieved through a process called "abliteration" using Heretic v1.4.0. This method involves targeted weight edits to the attention output and MLP down-projections, effectively suppressing refusal behavior without extensive fine-tuning.
Key Capabilities & Features
- Decensored Output: Significantly reduced refusal rates (from 52/100 to 8/100 adversarial prompts) compared to the original model, allowing it to comply with requests the base model would typically refuse.
- Retained Core Functionality: The underlying knowledge and instruction-following abilities of the Qwen2.5-Coder-0.5B-Instruct model, particularly its code-focused expertise, are largely preserved.
- Efficient Performance: With only 0.5 billion parameters and a 32768-token context length, it is designed for efficient local execution, including on CPU-only systems and Apple Silicon.
- Low KL Divergence: The abliteration process results in a low KL divergence of 0.1249 on the output distribution, indicating a narrow and targeted edit rather than a broad perturbation.
Ideal Use Cases
- Local Code Agents: Excellent for on-device code generation and assistance.
- Copilot-style Tools: Suitable for local, privacy-focused code completion and suggestion.
- Roleplay & Creative Applications: Useful for scenarios where uninhibited responses are desired.
- Research: Valuable for studying alignment, refusal mechanics, and the impact of targeted weight edits on LLM behavior.
Important Considerations
Users are responsible for the deployment and usage of this model, as it lacks safety filtering. It will comply with requests that the base model would refuse, including potentially harmful ones. The model inherits factual limitations and biases from its base, and abliteration does not enhance its inherent capabilities or judgment.