Justbackup/gemma-4-31B-it-uncensored
Justbackup/gemma-4-31B-it-uncensored is a 31 billion parameter instruction-tuned Gemma-4 model from Google, modified to significantly reduce refusal behavior. This model maintains the original quality while effectively removing content moderation filters, making it suitable for applications requiring less restrictive conversational outputs. It achieves a 3.2% refusal rate on cross-dataset validation, compared to 100% for the base model, without degradation in response quality.
Loading preview...
Overview
Justbackup/gemma-4-31B-it-uncensored is a 31 billion parameter instruction-tuned model based on Google's Gemma-4 architecture. Its primary distinction is the removal of refusal behavior, making it an "uncensored" variant of the original google/gemma-4-31B-it model. This modification was achieved using a novel norm-preserving biprojected obliteration method, which ensures the model's weight magnitudes are preserved, preventing quality degradation.
Key Capabilities
- Reduced Refusal Rate: Achieves a significantly lower refusal rate (3.2% on a 686-prompt cross-dataset validation) compared to the base model (100% refusal on 100 prompts).
- Quality Preservation: Maintains the original response quality, with a harmless response length ratio of approximately 1.01, indicating no degradation in output utility.
- Advanced Abliteration Method: Utilizes norm-preserving biprojected obliteration, which differs from standard projection by preserving weight magnitudes and employing per-layer refusal directions for more precise modification.
- Context Length: Supports a substantial context length of 32768 tokens.
Good For
- Applications requiring a large language model with minimal content moderation or refusal behaviors.
- Use cases where the base Gemma-4 model's safety filters are too restrictive.
- Developers interested in exploring the capabilities of a less constrained instruction-tuned model while retaining the underlying Gemma-4 architecture's performance.