RedactionBench is an evaluation-only benchmark for character-level redaction across eleven document categories. Each of the 200 documents is manually-annotated with character spans that are either mandatory (must redact) or contextual.
RedactionBench mixes 101 real-world documents manually sourced from the public web (transcribed, augmented) with 99 synthetic documents authored to fill categories where synthetic data is more appropriate.
