saidutta69/Qwen2.5-Coder-7B-Instruct-heretic
The saidutta69/Qwen2.5-Coder-7B-Instruct-heretic is a 7.6 billion parameter instruction-tuned causal language model, derived from Qwen/Qwen2.5-Coder-7B-Instruct. This variant has been modified using the Heretic v1.4.0 abliteration technique to suppress refusal behaviors, making it suitable for coding tasks without guardrails. It maintains the base model's knowledge and instruction-following capabilities while significantly reducing refusals, achieving 3/100 refusals compared to 100/100 for the base model.
Loading preview...
Overview
This model, Qwen2.5-Coder-7B-Instruct-heretic, is a specialized variant of the Qwen/Qwen2.5-Coder-7B-Instruct model, developed by saidutta69. It features 7.6 billion parameters and a 32K context length, primarily designed for code generation and assistance. Its key differentiator is the removal of refusal guardrails through a technique called abliteration (directional ablation) using Heretic v1.4.0.
Key Capabilities & Differentiators
- Decensored Coding Model: Suppresses refusal behaviors present in the base model, allowing it to answer requests that the original model would decline.
- Preserved Base Capabilities: Unlike fine-tuning, abliteration directly edits specific weights responsible for refusal, leaving the base model's core knowledge, instruction-following, and coding abilities largely intact.
- Low KL Divergence: Achieves an exceptionally low KL divergence of 0.0196 from the base model, indicating minimal impact on the overall output distribution while drastically reducing refusals.
- Performance: Reduced refusals from 100/100 to 3/100 on adversarial prompts, demonstrating effective decensoring without degrading coding performance.
Ideal Use Cases
- Local Coding Agents: Suitable for automated coding tasks where direct answers are preferred.
- Pair-Programming: Acts as an uninhibited assistant for developers.
- Repo-Level Assistance: Provides comprehensive code-related support without built-in content restrictions.
Important Considerations
Users are responsible for the deployment and use of this model, as it deliberately bypasses safety filters. It inherits the factual limitations and biases of the original Qwen2.5-Coder-7B-Instruct model.