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

Hugging Face
TEXT GENERATIONConcurrent Unit Cost:1Model Size:3.1BQuant:BF16Context Size:32kTool Calling:SupportedPublished:Mar 25, 2026License:qwen-researchArchitecture:Transformer0.0K Featherless Exclusive Warm

Qwen2.5-3B-Instruct-heretic is a 3.09 billion parameter instruction-tuned causal language model, a decensored variant of Qwen/Qwen2.5-3B-Instruct. Developed by saidutta69 using Heretic v1.2.0, it suppresses refusal behavior via targeted weight edits rather than fine-tuning, preserving the base model's knowledge and instruction-following. This model is optimized for use cases requiring direct answers without refusal, such as local agents, roleplay, or research into alignment mechanics.

Loading preview...

Overview

Qwen2.5-3B-Instruct-heretic is a specialized variant of the Qwen/Qwen2.5-3B-Instruct model, developed by saidutta69. This 3.09 billion parameter model has been "decensored" using the Heretic v1.2.0 method, which involves targeted weight edits to the attention output and MLP down-projections. This approach effectively suppresses refusal behavior without degrading the base model's core capabilities or coherence, unlike traditional fine-tuning methods.

Key Characteristics

  • Refusal Suppression: Achieves a significant reduction in refusals (2 out of 100 adversarial prompts) compared to the base model (96 refusals), as measured by Heretic's internal harness.
  • Preserved Capabilities: The underlying knowledge and instruction-following abilities of the base Qwen2.5-3B-Instruct model are largely maintained, as the edits are narrow and targeted.
  • Low KL Divergence: Exhibits a low KL divergence of 0.1327 from the base model's output distribution, indicating a precise and minimal perturbation.
  • Local Deployment: Designed for local deployment, supporting formats like GGUF (Q4_K_M, Q5_K_M) and compatible with llama.cpp and transformers.

Ideal Use Cases

  • Local Agents: Suitable for applications where a small, locally runnable model needs to provide direct answers.
  • Roleplay: Excels in scenarios requiring uninhibited conversational responses.
  • Alignment Research: Useful for studying refusal mechanics and alignment without the inherent guardrails of RLHF-era models.
  • Uncensored Applications: For any use case where over-refusal from standard instruction-tuned models is a hindrance.

Finetunes
1 models