PII-SCOPE
An open-source reference benchmark for PII detection, intended to address problems in publicly available LLM-generated datasets, including annotation noise, imbalanced distribution of PII types and insufficient contextual realism. By providing a standardized evaluation setting, it supports comparable assessment of different models on personally identifiable information detection tasks, allowing results across models to be measured on a common basis rather than on divergent, privately constructed data.
Timeline (last 90 days)
A 278M-parameter open-weights PII detection model has been released, along with the PII-SCOPE reference benchmark and a technical report.
The PII-SCOPE benchmark has also gone live on Hugging Face.
The motivation for building the PII-SCOPE benchmark was to address quality problems prevalent in public LLM-generated datasets (such as annotation noise, imbalanced PII type distribution, and insufficient contextual realism).
Currently available public information has not yet disclosed the specific scale, PII category coverage, or evaluation metric details of the PII-SCOPE benchmark.
All Facts (3)
The motivation for building the PII-SCOPE benchmark was to address quality problems prevalent in public LLM-generated datasets (such as annotation noise, imbalanced PII type distribution, and insufficient contextual realism).
50%UnverifiedA 278M-parameter open-weights PII detection model has been released, along with the PII-SCOPE reference benchmark and a technical report.
50%UnverifiedThe PII-SCOPE benchmark has also gone live on Hugging Face.
50%