A new open-weights model for detecting Personally Identifiable Information (PII) has been released, along with a reference benchmark and research paper. The model, a fine-tune of RoBERTa, was trained on 1.2 billion synthetic text samples containing various PII types. The accompanying benchmark, PII-SCOPE, is designed to overcome issues found in publicly available LLM-generated datasets and is available on Hugging Face. The model can identify PII instances by analyzing surrounding context and can either report the PII type and location or redact it from the input. AI
IMPACT Provides a new tool for identifying and redacting sensitive information in text, potentially improving data privacy in AI applications.
RANK_REASON Release of a new open-weights model, benchmark, and research paper. [lever_c_demoted from research: ic=1 ai=1.0]
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →
