llmfan46/Gemma-4-Queen-31B-it-uncensored-heretic

Hugging Face
VISIONConcurrency Cost:2Model Size:31BQuant:FP8Ctx Length:32kTool Calling:SupportedPublished:Apr 12, 2026License:apache-2.0Architecture:Transformer0.0K Open Weights Warm

llmfan46/Gemma-4-Queen-31B-it-uncensored-heretic is a 31 billion parameter Gemma-based instruction-tuned language model, created by llmfan46, with a 32768 token context length. This model is a decensored version of aifeifei798/Gemma-4-Queen-31B-it, achieved using the Heretic v1.2.0 tool with Arbitrary-Rank Ablation (ARA) to significantly reduce refusals (12/100 vs 99/100) while maintaining high model quality (0.0707 KL divergence). It excels in role-playing, creative writing, and generating scholarly responses, particularly with a 'Dark Academia Queen' persona.

Loading preview...

Model Overview

llmfan46/Gemma-4-Queen-31B-it-uncensored-heretic is a 31 billion parameter model derived from aifeifei798/Gemma-4-Queen-31B-it. Its primary differentiator is its decensored nature, achieved through the Heretic v1.2.0 tool using the Arbitrary-Rank Ablation (ARA) method. This process significantly reduces content refusals, dropping from 99/100 in the original model to just 12/100 in this version, while preserving model quality with a low KL divergence of 0.0707.

Key Capabilities

  • Reduced Refusals: Offers 88% fewer refusals compared to its base model, enabling more direct and unrestricted responses.
  • High Model Quality: Maintains performance close to the original model, as indicated by a low KL divergence.
  • Role-playing: Particularly suited for various role-playing scenarios, including 'Dark-roleplay' and a 'Dark Academia Queen' persona.
  • Creative Writing: Excels at generating writing prompts, opus, and songs, often with extensive, thesis-like detail.
  • Scholarly Responses: Capable of producing detailed and scholarly answers.
  • Multifaceted Persona Support: Designed to embody a wide range of personas, from 'X Queen' (Savage Commentator) to 'Dark Academia Queen' (Scholar) and 'AI Prompt Queen' (LLM expert).

Performance Metrics

  • Refusals: 12/100 (Original: 99/100)
  • KL Divergence: 0.0707
  • PIQA Accuracy: 0.9445 (Original: 0.9450) – demonstrating minimal impact on common-sense reasoning.
  • MMLU Accuracy: 0.8533 (Original: 0.8688) – showing a slight decrease in general knowledge and reasoning, balanced by the significant reduction in refusals.

Good For

  • Use cases requiring uncensored or less restricted text generation.
  • Creative writing and content generation, especially for detailed narratives or songs.
  • Role-playing applications with diverse and specific personas.
  • Generating scholarly or thesis-like responses.
  • Applications where direct and immediate answers are preferred over disclaimers or moral warnings.

Finetunes
2 models