Sabomako/Mistral-Small-3.2-24B-Instruct-2506-heretic
Sabomako/Mistral-Small-3.2-24B-Instruct-2506-heretic is a 24 billion parameter instruction-tuned language model, derived from unsloth/Mistral-Small-3.2-24B-Instruct-2506. This model has been decensored using the Heretic v1.3.0 tool with a custom ORBA obliteration method. It is specifically designed to significantly reduce refusal rates compared to its original counterpart, making it suitable for applications requiring less restrictive content generation.
Loading preview...
Overview
Sabomako/Mistral-Small-3.2-24B-Instruct-2506-heretic is a 24 billion parameter instruction-tuned model based on the unsloth/Mistral-Small-3.2-24B-Instruct-2506 architecture. It features a context length of 32768 tokens. The primary modification involves a decensoring process using the Heretic v1.3.0 tool, employing a custom ORBA obliteration method.
Key Capabilities
- Decensored Output: Significantly reduces content refusal rates compared to the original model.
- Instruction Following: Retains the instruction-following capabilities of the base Mistral-Small-3.2-24B-Instruct model.
- Reduced Refusals: Achieves a refusal rate of 1/100, a substantial decrease from the original model's 98/100.
Performance
This model demonstrates a KL divergence of 0.0583 when compared to the original model, indicating a controlled divergence from the base's output distribution while achieving its decensoring goal. The most notable performance difference is in its refusal rate, which is 1/100, in stark contrast to the original model's 98/100. This modification was achieved through refusal classification using an LLM judge (Gemma4-12B-it-Heretic).
Good For
- Applications where the original model's refusal rates are too high.
- Use cases requiring more permissive content generation.
- Developers seeking a less restrictive instruction-tuned model in the 24B parameter class.