mlabonne/gemma-3-1b-it-abliterated

Gemma 3 1B IT Abliterated: Uncensored Instruction-Tuned Model

This model is an uncensored variant of the google/gemma-3-1b-it instruction-tuned large language model, developed by mlabonne. It leverages a new "abliteration" technique to significantly reduce refusal behaviors often found in base models.

Key Capabilities & Features

• Uncensored Output: Modified to provide a high acceptance rate (>90%) for prompts that might otherwise be refused by standard instruction-tuned models.
• Abliteration Technique: Utilizes a layerwise abliteration method, computing a refusal direction based on hidden states across most layers (3 to 45) and applying a refusal weight of 0.75.
• Coherent Generation: Despite the experimental nature of the technique, the model aims to produce coherent outputs, though occasional minor text garbling (e.g., "It' my" instead of "It's my") may occur.
• Resilience: Gemma 3 architecture demonstrated higher resilience to this abliteration process compared to other models like Qwen 2.5.

Recommended Usage

• Generation Parameters: For optimal results, it is recommended to use temperature=1.0, top_k=64, and top_p=0.95.
• Quantization: GGUF quantizations are available for efficient deployment and inference.