Reza2kn/persian-pii-masking-openpii-690k-clean

Persian PII-Masking Combined Corpus, Cleaned

Cleaned Persian / Iranian PII-masking token-classification data.

This repo combines the audited persona-clean and initial-clean corpora. The combined split contains only rows kept after the source-specific audits.

• —Persona-clean rows: 623890
• —Initial-clean rows: 224956

Dataset Repo

Reza2kn/persian-pii-masking-openpii-690k-clean

Schema

Rows include:

• —source_text
• —masked_text
• —privacy_mask with label, start, end, value, label_index
• —mbert_tokens
• —mbert_token_classes
• —language
• —region
• —script
• —uid
• —split
• —seed_uid
• —persona_id (-1 for non-persona initial rows)
• —rendering_idx
• —gen_model

Splits

Cleaning

Cleaning was performed after all rows were embedded with the Mac Studio MLX Qwen3 embedding endpoint. The published clean artifacts remove exact hard failures and keep one representative per full-dataset nearest-neighbor component at cosine >= 0.95.

The local audit manifests used for this repo are uploaded under audit/ where applicable.

Synthetic Data

All personal data is synthetic and generated for PII detection/masking research. Values may be format-valid but do not represent real people.

Attribution

This dataset is localized from the English rows of `ai4privacy/pii-masking-openpii-1m`. License: CC-BY-4.0.