Ishowbackup/gemma-4-31B-it-uncensored
Ishowbackup/gemma-4-31B-it-uncensored is a 31 billion parameter instruction-tuned Gemma-4 model developed by Ishowbackup, featuring a 32K context length. This model has been specifically modified to remove refusal behavior, making it an 'uncensored' version of the original Google Gemma-4-31B-it. It is optimized for applications requiring direct answers without AI identity disclaimers or content refusals, while maintaining the original model's quality.
Loading preview...
Overview
Ishowbackup/gemma-4-31B-it-uncensored is a 31 billion parameter instruction-tuned model based on Google's Gemma-4-31B-it, with a 32K context length. Its primary distinction is the removal of refusal behavior, achieved through a novel "norm-preserving biprojected abliteration" method. This technique ensures that the model's weight magnitudes are preserved, preventing degradation in response quality while effectively eliminating content refusals.
Key Capabilities
- Reduced Refusal Behavior: Demonstrates a significant reduction in refusals, with only 1/100 effective refusals on mlabonne's 100 prompts dataset and 3.2% across a cross-dataset validation of 686 prompts.
- Quality Preservation: Maintains response quality, with a harmless response length ratio of approximately 1.01, indicating no degradation compared to the original model.
- Advanced Abliteration Method: Utilizes a norm-preserving biprojection technique that projects out refusal directions from weight components, ensuring
||W_new|| = ||W_orig||. - Efficient Processing: Employs per-layer refusal directions and a deterministic single-pass process, offering faster and equally effective results compared to other methods.
Good For
- Applications requiring direct, unfiltered responses from a large language model.
- Use cases where AI identity disclaimers or content refusals are undesirable.
- Developers seeking a powerful 31B parameter model with a high context length and modified behavioral characteristics.