InfinimindCreations/gemma-4-31B-it-uncensored
InfinimindCreations/gemma-4-31B-it-uncensored is a 30.7 billion parameter instruction-tuned model based on Google's Gemma 4 architecture, featuring a 32K context length. This variant has undergone norm-preserving biprojected abliteration to remove refusal behaviors, achieving 0.0% effective refusals across 656 adversarial prompts. It is optimized for responding to a broad range of queries without content restrictions, making it suitable for research into uncensored LLM responses.
Loading preview...
Overview
InfinimindCreations/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 version maintains awareness of risks and context, providing information rather than blocking queries.
Key Capabilities
- Zero Effective Refusals: Achieves 0.0% effective refusals across 656 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 biprojection-memeff mode, based on thegrimjimmethod, to remove refusal mechanisms while preserving model integrity. - Extensive Modification: 60 out of 60 layers (100%) and 120 weight matrices were modified to achieve uncensored behavior.
- Large Context Window: Supports a 32,768 token context length.
Good For
- Research into uncensored language model behavior and safety alignment.
- Applications requiring a model that will attempt to answer all prompts without built-in refusal mechanisms.
- Exploring the boundaries of LLM responses to sensitive or adversarial queries.