darkc0de/Mistral-Small-3.2-24B-Instruct-2506-Text-Only-heretic
darkc0de/Mistral-Small-3.2-24B-Instruct-2506-Text-Only-heretic is a 24 billion parameter instruction-tuned causal language model based on the Mistral architecture. This model is a decensored version of anthracite-core/Mistral-Small-3.2-24B-Instruct-2506-Text-Only, created using the Heretic tool. It features a standard Mistral architecture without a vision encoder and is specifically optimized to reduce refusals compared to its original counterpart, making it suitable for applications requiring less restrictive content generation.
Loading preview...
Model Overview
This model, darkc0de/Mistral-Small-3.2-24B-Instruct-2506-Text-Only-heretic, is a 24 billion parameter instruction-tuned language model. It is a modified version of the anthracite-core/Mistral-Small-3.2-24B-Instruct-2506-Text-Only model, specifically engineered for decensorship using the Heretic v1.1.0 tool.
Key Characteristics
- Decensored Nature: The primary differentiator of this model is its significantly reduced refusal rate. While the original model had a refusal rate of 98/100, this 'heretic' version demonstrates a refusal rate of only 4/100, indicating a much less restrictive output generation.
- Architecture: It utilizes a standard "Mistral" architecture and does not include a vision encoder, making it a text-only model.
- Abliteration Parameters: Specific parameters were adjusted during the decensoring process, including
attn.o_projandmlp.down_projweights, to achieve the desired reduction in content filtering.
Performance
Compared to the original model, this version shows a KL divergence of 0.1604, reflecting the changes introduced during the decensoring process.
Use Cases
This model is particularly suited for applications where a less constrained and more direct response generation is desired, especially in scenarios where the original model's high refusal rate would be a limitation.