sci4ai/Qwen2.5-Coder-14B-Abliterated

Hugging Face
TEXT GENERATIONConcurrent Unit Cost:1Model Size:14.8BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Mar 29, 2026License:apache-2.0Architecture:Transformer Open Weights Featherless Exclusive Warm

Qwen2.5-Coder-14B-Abliterated is a 14.8 billion parameter language model developed by sci4ai, based on the Qwen2.5-Coder-14B-Instruct architecture with a 32768 token context length. This model has undergone 'abliteration' to remove refusal behaviors, specifically targeting code-related safety guardrails. It is optimized for code generation tasks, providing responses that the original model would typically refuse.

Loading preview...

Qwen2.5-Coder-14B-Abliterated: Refusal-Removed Code Generation

This model is an 'abliterated' version of the Qwen/Qwen2.5-Coder-14B-Instruct, a 14.8 billion parameter model with a 32768 token context length. The primary modification involves the removal of refusal behaviors, particularly those related to generating harmful or unethical code, through a technique called 'abliteration'.

Key Capabilities

  • Unrestricted Code Generation: Designed to provide code outputs for requests that the base Qwen2.5-Coder model would typically refuse due to safety guardrails.
  • Activation-Based Weight Surgery: Achieves refusal removal by identifying and ablating 'refusal directions' within the model's o_proj and down_proj weight matrices across 47 of its 48 layers.
  • Research-Oriented: Intended for research purposes to explore the effects of removing safety mechanisms in large language models.

Good For

  • Exploring Model Limitations: Researchers interested in studying model safety, refusal mechanisms, and the impact of their removal.
  • Unfiltered Code Generation: Use cases where the original model's safety filters are undesirable, such as red-teaming or specific research scenarios requiring potentially harmful code outputs.

Note: This model is provided for research and users are responsible for its deployment and outputs, as it will comply with requests the original model would refuse.