saidutta69/Qwen2.5-3B-Instruct-heretic

Overview

This model, saidutta69/Qwen2.5-3B-Instruct-heretic, is a 3.09 billion parameter instruction-tuned causal language model created by RACER IS OP. It is a decensored variant of the original Qwen/Qwen2.5-3B-Instruct, achieved through a process called "abliteration" using Heretic v1.2.0. This method involves targeted weight edits to the attention output and MLP down-projections to suppress refusal behavior, rather than fine-tuning, thereby preserving the base model's core knowledge and instruction-following abilities.

Key Differentiators

• Decensored Output: Significantly reduces refusal behavior, with only 2 refusals out of 100 adversarial prompts compared to 96 for the base model.
• Targeted Modification: Utilizes abliteration to precisely edit weights responsible for refusal, minimizing impact on overall model coherence and capabilities.
• Low KL Divergence: Exhibits a low KL divergence of 0.1327 from the base model, indicating a narrow and targeted modification rather than a broad perturbation.

Good For

• Local Agents: Ideal for applications where a small, locally runnable model needs to provide direct answers without censorship.
• Roleplay: Suitable for scenarios requiring the model to engage in roleplay without refusing certain prompts.
• Research: Useful for studying alignment, refusal mechanics, or the effects of targeted weight edits on LLM behavior.
• Use Cases Blocked by RLHF: Addresses scenarios where the over-refusal common in RLHF-era models is a hindrance.