SAEI-CLINDDx validation
The schema.json file uses JSON Schema Draft 2020-12 and describes one JSONL case record (not an array). The files data/es/test.jsonl and data/en/test.jsonl share exactly the same fields and case IDs. Both languages remain testing-only.
Run
python -m pip install 'jsonschema>=4.18'
python scripts/validate_dataset.py --report reports/validation_results.json
python scripts/validate_dataset.py --schema-only
python scripts/validate_dataset.py --expected-count 251
Exit code 0 means no errors; 1 means validation errors; 2 means the schema cannot be loaded. Warnings do not cause failure. JSON reports contain file path, JSONL line number, case ID, field, code, severity and message.
Checks
- JSON Schema: required fields, string/integer types, valid language, ID pattern, core nonempty clinical text, no unknown keys.
- Dataset integrity: parse errors, blank JSONL lines, duplicate IDs, ID/year consistency, correct language directory, bilingual ID set equality, aligned year and author strings.
- Content QA (default): flags spreadsheet error strings such as
#ERROR!in any clinical text as errors and empty differential-diagnosis fields as warnings. Content QA is a screening step, not full MQM evaluation or a clinical quality guarantee.
summary, differential_diagnosis, authors, and discussion permit empty strings to preserve missingness in the original dataset. These are not replaced with invented content. The optional --expected-count flag is deliberately absent by default so new cases can be added without editing the validator.
Known status of current English version
The current English data contains 13 #ERROR! final-diagnosis values. These retain string type and therefore pass JSON Schema, but fail content QA. The files are intentionally left unchanged for traceability. Resolve those failures, then re-run the complete checks before release. No structural validator can guarantee accurate medical translation; use a separate human-adjudicated MQM evaluation.