saidutta69/Qwen2.5-Coder-3B-Instruct-heretic
The saidutta69/Qwen2.5-Coder-3B-Instruct-heretic is a 3.1 billion parameter instruction-tuned causal language model, derived from Qwen's Qwen2.5-Coder-3B-Instruct. This variant has been modified using the Heretic v1.2.0 tool to suppress refusal behaviors through targeted weight edits, rather than fine-tuning. It is specifically designed for developers seeking a code-focused model without built-in refusal guardrails, excelling in local coding agents, copilot-style assistance, and direct code generation. The model maintains the base Qwen model's knowledge and instruction-following capabilities while offering enhanced compliance for coding tasks.
Loading preview...
Qwen2.5-Coder-3B-Instruct-heretic Overview
This model is a specialized 3.1 billion parameter variant of the Qwen2.5-Coder-3B-Instruct, developed by saidutta69. Its core distinction lies in its "decensored" nature, achieved through a process called "abliteration" using the Heretic v1.2.0 tool. This method involves targeted weight edits to the attention output and MLP down-projections to suppress refusal behaviors, allowing the model to comply with requests that the original base model would typically refuse.
Key Capabilities & Features
- Refusal Suppression: Significantly reduces refusal rates (3/100 compared to 100/100 for the original model) without compromising the base model's knowledge or instruction-following abilities.
- Code-Focused: Inherits the strong code generation and understanding capabilities of the Qwen2.5-Coder-3B-Instruct base model.
- Resource Efficient: At 3.1B parameters, it is designed to run efficiently on consumer hardware, including gaming PCs and CPUs, with various GGUF quantizations available.
- Direct Code Generation: Optimized for scenarios requiring direct answers and code output without built-in safety guardrails.
Ideal Use Cases
- Local Coding Agents: Suitable for developers building local AI assistants for coding tasks.
- Copilot-Style Assistance: Provides direct, uninhibited code suggestions and completions.
- Unfiltered Code Generation: For applications where the primary goal is to generate code or technical solutions without refusal behaviors.
It's important to note that while refusal suppression is deliberate, users are responsible for its deployment, as it lacks additional safety filtering. The model inherits the factual limitations and biases of its base model.