What is Gemma 3 27B IT QAT Abliterated?

This model is an uncensored version of Google's Gemma 3 27B instruction-tuned model, specifically google/gemma-3-27b-it-qat-q4_0-unquantized. Developed by mlabonne, it utilizes a unique abliteration technique to significantly reduce refusal behaviors present in the base model.

Key Capabilities

• Refusal Mitigation: Employs a novel abliteration method that computes and subtracts a 'refusal direction' from the model's hidden states, particularly targeting modules like o_proj.
• Enhanced Acceptance Rate: Achieves an acceptance rate exceeding 90% on a dedicated test set, evaluated using both a dictionary approach and NousResearch/Minos-v1.
• Coherent Output Generation: Designed to produce coherent and relevant responses while minimizing content restrictions.

Good For

• Applications requiring a less restrictive language model.
• Use cases where the base Gemma 3 model's refusal behaviors are undesirable.
• Exploration of advanced uncensoring techniques in large language models.