llmfan46/gemma-4-31B-it-uncensored-heretic
llmfan46/gemma-4-31B-it-uncensored-heretic is a 31 billion parameter instruction-tuned Gemma 4 model, developed by llmfan46, that has been decensored using the Heretic v1.2.0 Arbitrary-Rank Ablation (ARA) method. It achieves 90% fewer refusals (10/100) compared to the original Google Gemma-4-31B-it model, while preserving model quality with a KL divergence of 0.0541. This model is optimized for use cases requiring less content restriction without significant performance degradation.
Loading preview...
Overview
This model, llmfan46/gemma-4-31B-it-uncensored-heretic, is a 31 billion parameter instruction-tuned variant of Google's Gemma 4 model. It has been specifically modified using the Heretic v1.2.0 Arbitrary-Rank Ablation (ARA) method to significantly reduce content refusals. The modification results in 90% fewer refusals (10/100) compared to the original model (99/100 refusals), while maintaining a low KL divergence of 0.0541, indicating strong preservation of the original model's quality and capabilities.
Key Capabilities & Performance
- Decensored Output: Achieves a substantial reduction in content refusals, making it suitable for applications requiring less restrictive content generation.
- Multimodal: Inherits Gemma 4's capabilities for processing text and image inputs, with a context window of up to 256K tokens.
- Reasoning & Coding: Designed for strong reasoning, agentic workflows, and enhanced coding capabilities, including native function-calling support.
- MMLU Performance: Maintains a high MMLU accuracy of 85.90%, closely matching the original model's 86.50%.
Good for
- Developers seeking a powerful 31B parameter model with significantly reduced content moderation for broader application use cases.
- Applications where the original Gemma 4's refusal rate was a limiting factor.
- Tasks requiring robust reasoning, coding, and multimodal understanding with a preference for less constrained output.