OpenPII Text Entries with 27 Sensitive Data Classes

Comprising 220,000 text entries for training models to mask personally identifiable information (PII). It includes 27 distinct PII classes and targets 749 discussion subjects across education, health, and psychology domains. The dataset is authored by ai4privacy and was last updated in April 2024.

Use Cases

• Train a sequence labeling model to detect and mask the 27 PII classes within the 220k text entries.
• Evaluate model performance on PII masking across 749 distinct discussion subjects in education, health, and psychology contexts.
• Analyze the distribution and co-occurrence of different PII classes within the text entries.
• Fine-tune a language model for privacy-preserving text generation using the PII-labeled examples.

Strengths

• Contains 220,000 text entries, providing a substantial corpus for model training.
• Covers 27 distinct classes of personally identifiable information.
• Targets 749 specific discussion subjects across three application domains.
• Dataset is explicitly designed for privacy masking, a critical task for AI assistants and LLMs.

Limitations

• The total dataset size and specific file formats are unknown, complicating download and storage planning.
• The column structure and sample data are unavailable, making initial data exploration difficult.
• Potential bias towards the three specified domains (education, health, psychology) may limit generalizability to other contexts.

Provenance

Source

huggingface, author ai4privacy

Collection Method

null

Time Range

null

Freshness

Last updated on 2024-04-21.

Geography

null