treadon/

abliteration-eval

A balanced evaluation dataset for testing abliterated (uncensored) language models. Measures both refusal removal (does the model stop refusing harmful prompts?) and over-refusal correction (does the model stop incorrectly refusing safe prompts?).

	Dataset Structure




	
	Splits




	

Split Rows Purpose

harmful 200 Prompts that safety-trained models refuse. An abliterated model should comply.

over_refusal 83 Safe prompts that contain trigger… See the full description on the dataset page: https://huggingface.co/datasets/treadon/abliteration-eval.