SAHA-AL is a benchmark for training and evaluating text anonymization systems. It goes beyond detection accuracy by evaluating anonymization as a system under attack — measuring adversarial re-identification risk, contextual privacy leakage, and a formalized privacy-utility tradeoff.
3 evaluation tasks: PII detection, text anonymization quality, and adversarial privacy risk
11 metrics spanning leakage, utility, format… See the full description on the dataset page:
https://huggingface.co/datasets/huggingbahl21/saha-al.
