Ibrahimbaroud/roberta-large-ipi-mimiciii

Model Description

• A FacebookAI/roberta-large model fine-tuned for the detection of Indirect Personal Identifiers (IPIs) in medical notes.
• The model was trained to predict IPI spans following the schema described in Baroud et al., 2025, https://aclanthology.org/2025.privatenlp-main.7/.
• A token can either be classified as non-IPI or as one of the IPI types. Token predictions are aggregated to spans by making use of BILOU tagging.

Dataset

• The IPI annotations on MIMIC-III that were used for the training and the annotation guidelines can be found here: https://zenodo.org/records/15372705