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

TEXT GENERATIONConcurrent Unit Cost:1Model Size:0.5BQuant:BF16Context Size:32kTool Calling:SupportedPublished:Jul 16, 2026License:apache-2.0Architecture:Transformer Open Weights Featherless Exclusive Cold

The saidutta69/Qwen2.5-Coder-0.5B-Instruct-heretic model is a 0.5 billion parameter instruction-tuned variant of Qwen/Qwen2.5-Coder-0.5B-Instruct, developed by RACER IS OP. This model has been decensored using Heretic v1.4.0, suppressing refusal behaviors through targeted weight edits rather than fine-tuning. It is optimized for code-focused tasks, offering a small, efficient solution for local code agents and copilot-style use without typical refusal guardrails.

Loading preview...

Overview

saidutta69/Qwen2.5-Coder-0.5B-Instruct-heretic is a specialized variant of the Qwen/Qwen2.5-Coder-0.5B-Instruct model, developed by RACER IS OP. This 0.5 billion parameter model has undergone a "decensoring" process using the Heretic v1.4.0 tool, which employs directional ablation to suppress refusal behaviors. Unlike traditional fine-tuning, this method involves targeted weight edits to the attention output and MLP down-projections, ensuring that the base model's core knowledge and instruction-following capabilities remain largely intact.

Key Capabilities & Differentiators

  • Decensored Behavior: Significantly reduces refusal rates (from 52/100 to 8/100 adversarial prompts) compared to the original model, allowing it to comply with requests that the base model would typically refuse.
  • Code-Focused: Inherits the code-centric capabilities of the Qwen2.5-Coder series, making it suitable for programming-related tasks.
  • Efficient Deployment: With 0.5 billion parameters, it is designed to run comfortably on CPU and is ideal for on-device deployment and local applications.
  • Minimal Knowledge Perturbation: Achieves a low KL divergence of 0.1249, indicating that the modifications are narrow and targeted, preserving the original model's output distribution and capabilities.

Ideal Use Cases

This model is particularly well-suited for:

  • Local Code Agents: Powering automated coding assistants or tools that require high compliance.
  • Local Copilot-style Applications: Providing code suggestions and completions without encountering refusal guardrails.
  • Research on Alignment/Refusal Mechanics: Studying and experimenting with how refusal behaviors are implemented and suppressed in LLMs.
  • Any Use Case Blocked by Over-Refusal: Scenarios where the strict safety filters of RLHF-era models hinder desired functionality, especially in code generation or creative tasks.

Finetunes
1 models