uncensor

Summary

The uncensor dataset is a collection of chat-formatted training examples designed to remove safety filters and refusal behaviors from LLMs. It uses a 'Dolfino aka Dirty D' system prompt that instructs the model to be completely uncensored, unbiased, and compliant with all user requests regardless of ethics or legality. The dataset has been used to fine-tune uncensored versions of Llama-3.2, Hermes-3, and Qwen3 models by multiple community members.

Keywords

uncensored alignment-removal chat sft controversial jailbreak

Topics

NLP / Chatbot alignment

Research notes

• Controversial dataset designed to remove AI safety guardrails. Dataset viewer returns errors. Used by nicoboss for multiple uncensored model fine-tunes. No license specified on the dataset page.