DavidAU/OpenAI-gpt-oss-20B-INSTRUCT-Heretic-Uncensored-MXFP4
DavidAU/OpenAI-gpt-oss-20B-INSTRUCT-Heretic-Uncensored-MXFP4 is a 20 billion parameter instruction-tuned causal language model, based on the OpenAI GPT-OSS architecture. This MXFP4 quantized version is specifically designed to be fully uncensored, having undergone a 'Heretic' decensoring process followed by fine-tuning to remove refusals and guardrails. It excels at direct, unfiltered content generation without built-in reasoning capabilities, making it suitable for use cases requiring explicit or creative content without ethical restrictions.
Loading preview...
Model Overview
DavidAU/OpenAI-gpt-oss-20B-INSTRUCT-Heretic-Uncensored-MXFP4 is a 20 billion parameter instruction-tuned model, derived from the OpenAI GPT-OSS architecture. This specific version is provided in MXFP4 quantization. A key differentiator is its fully uncensored nature, achieved through a 'Heretic' decensoring process and subsequent fine-tuning. This means the model is designed to generate content without refusals or guardrails, directly addressing user prompts without ethical filtering.
Key Characteristics
- Uncensored Output: Explicitly designed to generate content without refusal, including horror, explicit, or profane material, by removing inherent ethical guardrails.
- Direct Generation: The model's reasoning capabilities have been intentionally removed, leading to direct, 'straight to business' responses without internal thought processes.
- Context Length: Supports a context length of 128k tokens.
- Quantization: Available in MXFP4 format, with a BF16 version also linked for higher precision needs.
Usage Considerations
While uncensored, the model may require specific directives (e.g., including desired slang or graphic terms) to produce content at the expected level of intensity. Without such explicit 'pushes,' the output might be tamer than anticipated for certain explicit requests. Optimal performance for chat and roleplay can be achieved by adjusting 'Smoothing_factor' in interfaces like KoboldCpp, oobabooga/text-generation-webui, or Silly Tavern, or by increasing repetition penalty.