RedacBench is a comprehensive benchmark for evaluating policy-conditioned redaction — the selective removal of sensitive information from text according to a given security policy. Unlike existing benchmarks that focus on predefined categories such as PII or on specific techniques like masking, RedacBench evaluates whether sensitive information remains inferable after redaction under policy-defined constraints, across individual, corporate… See the full description on the dataset page:
https://huggingface.co/datasets/hyunjunian/RedacBench.
