Privasis-Enterprise-Train (v3, 37K)

The training partition for need-to-know document sanitization, generated by the same v3 contextual pipeline as Privasis-Enterprise-Eval (v3). Company names are disjoint from the eval split (unseen-entity generalization).

Each record is a supervised sanitization example:

(records, sharing instruction) -> sanitized records

The model is given a bundle of internal company documents about one business event and a natural-language sharing instruction (what a specific recipient, for a specific purpose, may and may not receive), and must return the bundle rewritten so protected facts are removed or abstracted and needed facts are kept. Two instruction tracks are provided per record: explicit (names the categories to remove/keep) and context-only (recipient + purpose only).

Size

One JSON object per line (train.jsonl); one line per (event × sharing-context) instance.

	count
context records	41,317
business-event cases	22,490
sharing contexts	21,325 events × 2

Same schema as the v3 eval card (fields: documents [text, gold_sanitized], instruction, instruction_context_only, key_infos, composite_infos, targets, retention, gold_verbatim_leak_ids, cross_record_verified). The reference gold_sanitized is the teacher's (gpt-oss-120b) chunk-level sanitization; a protected value survives verbatim in only ~0.25% of targets (each flagged in gold_verbatim_leak_ids).

Generation

Event-first synthesis → per-event sharing-context pair with need-to-know boundaries (PROTECT/RETAIN/OPTIONAL, sensitivity as prior only) → chunk-level reference sanitization with a targeted context-preserving abstraction pass. No evaluation stage (training data). All entities invented; no real personal data.

License

CC-BY-NC-4.0. Synthetic data for privacy research and development. Not for commercial use.