Iambackup/gemma-4-31B-it-uncensored
Iambackup/gemma-4-31B-it-uncensored is a 30.7 billion parameter instruction-tuned causal language model based on Google's Gemma 4 architecture, featuring a 32K context length. This model has undergone norm-preserving biprojected abliteration to remove refusal behaviors, ensuring it responds to all prompts without censorship. It is specifically designed for applications requiring an uncensored large language model that engages with all topics, providing factual information rather than blocking queries.
Loading preview...
Overview
Iambackup/gemma-4-31B-it-uncensored is a 30.7 billion parameter instruction-tuned model derived from Google's Gemma 4 31B-it. Its primary distinction is the complete removal of refusal behaviors through a process called norm-preserving biprojected abliteration, ensuring the model responds to all prompts. This uncensored variant maintains a 32K context length and is the largest dense model in the Gemma 4 family available in this configuration.
Key Capabilities
- Zero Effective Refusals: Achieves 0.0% effective refusals across 656 diverse prompts from datasets like JailbreakBench, forbidden_questions, and BeaverTails. Automated flagging identified 14 responses, but manual audit confirmed these were false positives (e.g., "As an AI, I don't have a physical body" to sexual requests, or medical disclaimers).
- Norm-Preserving Abliteration: Utilizes the
heretictool with a biprojection-memeff method, applying norm-preserving orthogonalized ablation (grimjim) to 100% of the model's layers. This method ensures the removal of refusal mechanisms while preserving model integrity. - Comprehensive Modification: 60 out of 60 layers and 120 weight matrices were modified, including specific patches for Gemma 4's
Gemma4ClippableLinearto ensure full uncensoring.
Good For
- Research into Uncensored LLMs: Ideal for researchers studying the behavior and implications of large language models without built-in refusal mechanisms.
- Applications Requiring Full Engagement: Suitable for use cases where the model must engage with all topics and provide information, even on subjects typically flagged by safety filters.
- Exploring Model Limitations: Useful for understanding how models respond when their inherent safety guardrails are removed, providing factual information rather than blocking sensitive queries.