llmfan46/gemma-4-31B-it-qat-q4_0-unquantized-uncensored-heretic
llmfan46/gemma-4-31B-it-qat-q4_0-unquantized-uncensored-heretic is a 31 billion parameter instruction-tuned Gemma 4 model, developed by llmfan46, based on Google DeepMind's architecture. This version is specifically decensored using the Heretic v1.2.0 tool with Arbitrary-Rank Ablation (ARA) to significantly reduce refusals while preserving model quality. It maintains a 32768 token context length and is optimized for applications requiring less restrictive content generation.
Loading preview...
Model Overview
This model, llmfan46/gemma-4-31B-it-qat-q4_0-unquantized-uncensored-heretic, is a 31 billion parameter instruction-tuned variant of Google DeepMind's Gemma 4 model. It has been decensored using the Heretic v1.2.0 tool with the Arbitrary-Rank Ablation (ARA) method, specifically targeting the attn.o_proj components.
Key Differentiators
- Reduced Refusals: Achieves an 89% reduction in refusals (11/100 vs. 99/100 for the original model) while maintaining a low KL divergence of 0.0365, indicating strong preservation of original model quality.
- Gemma 4 Foundation: Inherits core capabilities from the Gemma 4 family, including a 256K token context window, multimodal support (text and image), and strong reasoning and coding abilities.
- Quantization-Aware Training (QAT): Based on a QAT-optimized checkpoint, allowing for similar quality to bfloat16 with reduced memory requirements.
Performance
While significantly reducing refusals, the model shows a slight decrease in MMLU accuracy (84.46% for Heretic vs. 86.17% for the original), demonstrating a trade-off for increased uncensored output.
Good for
- Use cases requiring less restrictive content generation and fewer refusals.
- Applications benefiting from the Gemma 4 architecture's multimodal capabilities and long context handling.
- Developers seeking a powerful, uncensored 31B parameter model for text and image tasks.