What the fuck is this model about?
This model, arnomatic/Mistral-Small-3.2-24B-Instruct-2506-Text-Only-heretic, is a 24 billion parameter instruction-tuned language model derived from the Mistral-Small-3.2-24B-Instruct-2506 base. Its core purpose is to provide a decensored version of the original model, achieved through a process called "abliteration" using the Heretic v1.1.0 tool.
What makes THIS different from all the other models?
The primary differentiator of this model is its significantly reduced refusal rate compared to its base model. While the original model had a refusal rate of 97 out of 100 test cases, this "heretic" version demonstrates only 4 refusals out of 100. This makes it distinct from standard instruction-tuned models that often incorporate strong safety and refusal mechanisms. It maintains the standard Mistral architecture and does not include a vision encoder.
Abliteration Parameters
The decensoring process involved specific adjustments to attention and MLP weights, including:
attn.o_proj.max_weight: 1.17mlp.down_proj.max_weight: 1.27
Should I use this for my use case?
This model is particularly suited for use cases where the reduction of content refusals and censorship is a critical requirement. If your application demands a language model that is less prone to filtering or declining to answer certain prompts, this "heretic" variant offers a distinct advantage. Developers should consider this model if their use case involves generating content that might be restricted by more heavily moderated LLMs, while still benefiting from the underlying Mistral-Small-3.2 architecture.