UnsafeChain is a correction-based alignment dataset designed to improve the safety of large language models (LLMs) through exposure to unsafe completions and their corrected responses.
It contains three distinct subsets:
train_full.csv: Complete dataset with all examples.
train_random.csv: Randomly selected examples for baseline comparisons.
train_selected.csv: Carefully curated hard unsafe examples, emphasizing safety recovery.
Dataset Details… See the full description on the dataset page: https://huggingface.co/datasets/raj-tomar001/UnSafeChain.
