Raymond102103028/PATCH

PATCH: Prompt Assortment for Traditional Chinese Hazards The first large-scale adversarial safety dataset for Traditional Chinese (TC), designed to train and evaluate content safety classifiers for lightweight LLMs. For full documentation, see our GitHub repository. Dataset Overview 593,020 safe prompts localized to Traditional Chinese 231,924 unsafe prompts across 13 MLCommons hazard categories PATCH-GPT: Direct harmful prompts PATCH-RT: Evasive prompts… See the full description on the dataset page: https://huggingface.co/datasets/Raymond102103028/PATCH.