575-lab/kiji-inspector-reviewed-pairs

Kiji PII Detection Training Data

Synthetic multilingual dataset for training PII (Personally Identifiable Information) detection models with token-level entity annotations and coreference resolution.

Dataset Summary

Usage

from datasets import load_dataset

ds = load_dataset("575-lab/kiji-inspector-reviewed-pairs")

# Access a sample
sample = ds["train"][0]
print(sample["text"])
print(sample["privacy_mask"])   # PII entity annotations
print(sample["coreferences"])   # Coreference clusters
print(sample["language"])       # e.g. "English"
print(sample["country"])        # e.g. "United States"

Schema

Each sample contains:

Example sample

{
  "text": "Contact Dr. Maria Santos at maria.santos@hospital.org or call +1-555-123-4567.",
  "privacy_mask": [
    {"value": "Maria", "label": "FIRSTNAME"},
    {"value": "Santos", "label": "SURNAME"},
    {"value": "maria.santos@hospital.org", "label": "EMAIL"},
    {"value": "+1-555-123-4567", "label": "PHONENUMBER"}
  ],
  "coreferences": [
    {
      "mentions": ["Dr. Maria Santos", "maria.santos"],
      "entity_type": "FIRSTNAME",
      "cluster_id": 0
    }
  ],
  "language": "English",
  "country": "United States"
}

PII Labels

Language Distribution

Country Distribution

Data Generation

Samples are synthetically generated using LLMs with structured outputs. The generation pipeline:

1. 1.NER generation — LLM produces text with embedded PII and entity annotations
2. 2.Coreference generation — second pass links pronouns and references to their antecedent entities
3. 3.Review (optional) — additional LLM pass validates and corrects annotations
4. 4.Format conversion — samples are converted to a clean, standardized schema

Intended Use

This dataset is designed for training token-classification models that detect and classify PII in text. The coreference annotations enable training models that can also resolve entity mentions (e.g., linking "he" back to "John Smith").

Limitations

• —All data is synthetically generated — entity distributions may not match real-world text
• —Coreference annotations are LLM-generated and may contain errors
• —Address and ID formats are country-specific but may not cover all regional variations