Uncensored-SFT-v2 (High Quality Uncensored Instruction Dataset V2)

Summary

A semantically deduplicated instruction-tuning dataset of ~485k input/output pairs covering uncensored, harmful, and adversarial prompts (scams, hacking, explicit content, etc.) with compliant responses. It is intended for training 'uncensored' LLMs that do not refuse requests and for safety/alignment research on model behavior. V2 is a deduplicated version of the original V1 release.

Keywords

uncensored sft instruction-tuning harmful-prompts adversarial red-teaming safety-research alignment

Topics

NLP / Instruction Tuning

Research notes

• Contains explicit, harmful, and illegal-activity content. Intended for research purposes only; not for deployment without safety considerations.