dataset-redaction

To redact PHI in health datasets for benchmarking, apply deterministic pseudonymization to patient identifiers across visits. This removes sensitive information while preserving the document structure necessary for accurate OCR, STT, and LLM testing.

Synthetic data generation for benchmarking produces realistic, test-ready data with controlled deltas to evaluate model performance. You need it when creating safe evaluation datasets that require realistic variations without exposing actual patient health information.

Yes, you can preserve cross-visit timing structures when redacting patient identifiers. The redaction pipeline enforces deterministic pseudonymization specifically designed to maintain these temporal relationships across visits while ensuring health datasets remain safe.

To generate redacted_documents.json and gold_facts.json for evaluation pipelines, run the redaction pipeline on your sample dataset. After generating these artifacts, validate the dataset using the provided manifests to ensure reproducible benchmarking.

Deliverables needed for reproducible dataset redaction and benchmarking include redaction policies, schemas, and tooling specifications. These ready-to-use artifacts ensure your evaluation data pipelines remain transparent and reproducible across different testing environments.

Yes, this redaction approach works for both STT and LLM test corpora. It applies deterministic pseudonymization to redact PHI and generate synthetic equivalents across health datasets, preserving the audio and text document structures required for benchmarking.