mlabonne/gemma-3-12b-it-abliterated

Overview

mlabonne/gemma-3-12b-it-abliterated is an uncensored version of the google/gemma-3-12b-it model, developed by mlabonne. This 12 billion parameter model has been modified using an experimental "abliteration" technique to reduce refusals and generate less restricted content. The developer notes that Gemma 3 models demonstrated higher resilience to this technique compared to other architectures like Qwen 2.5.

Key Characteristics

• Uncensored Output: Modified to remove refusal behaviors, aiming for a high acceptance rate (>90%) for various prompts.
• Abliteration Technique: Employs a layerwise abliteration method, computing a refusal direction based on hidden states across most layers (3 to 45) and applying a refusal weight of 0.6.
• Experimental Nature: The technique is experimental, and occasional garbled text (e.g., "It' my" instead of "It's my") has been observed.
• Recommended Generation Parameters: Users are advised to use temperature=1.0, top_k=64, and top_p=0.95 for optimal results.

Use Cases

This model is suitable for applications where less restrictive content generation is desired, particularly in scenarios where the base Gemma 3 model might exhibit excessive refusal behaviors. Its uncensored nature makes it potentially useful for creative writing, role-playing, or research into model safety and bias mitigation.