// 01 / Method
How abliteration works
Researchers identified an internal direction associated with refusing prompts in the models they studied. Model builders can suppress that direction by editing model weights, which can make the model less likely to refuse certain prompts.
“Uncensored” is a broader description. A provider might use it for an abliterated model, a differently fine-tuned model, or its own approach to filtering. Look at the model documentation and the API terms separately.
Read the refusal-direction research · Hugging Face explainer
// 02 / Checklist
What to check before you integrate
- Confirm that the exact model ID is available through the provider.
- Read the provider’s current acceptable-use terms for your application.
- Compare input and output rates for a representative request.
- Test representative prompts and record the model and endpoint you used.
// 03 / Providers
Compare hosted options
Refuseless and Unaligned publish abliterated model options. Compare their model lists, compatibility, and payments in the text directory.