Clinical NLI checkpoint preparation¶

This optional research recipe prepares #3236 without a library-release milestone. Training and publishing a new clinical NLI checkpoint are not prerequisites for feature APIs or SDK releases. Existing independently qualified local models or caller-supplied calibrated providers can satisfy the runtime contract. Missing or unqualified providers still fail closed. No checkpoint, registry alias, or quality result is released by this document.

The configuration at openmed/training/configs/clinical_nli_small.yaml starts from the Apache-2.0 ClinicalE5-Small encoder at immutable revision 2ea947130eb28d610ceb96472af0b694fce9c344. Its token-classification head must be discarded and replaced with a three-class sequence head. The configured seed, split, optimizer, and export formats are proposed starting values, not evidence of a training run.

Before any training, record a public-data manifest with dataset URLs, immutable revisions, license review, content hashes, duplicate checks, and source-grouped splits. Synthetic examples must be clearly tagged. MedNLI is user-supplied and evaluation-only; no MedNLI data or derived examples may enter training, the repository, or a published artifact. Patient data is excluded.

A candidate must report public and synthetic three-class accuracy separately, negation false-entailment rate, error slices, and calibrated abstention. The PyTorch, ONNX int8, and MLX exports need pinned checksums, offline load tests, and prediction parity. BenchmarkReport and gate results must contain only counts, rates, hashes, model IDs, and revision IDs. The model card must explain the three-way-plus-abstain contract and evidence limits. A MedNLI number, if the owner has lawful local access, is one evaluation-only aggregate line.

openmed.eval.nli_gate.build_nli_preparation_report composes the existing error-slice, calibration, and negation reports with public and synthetic accuracy counts. Its BenchmarkReport retains digests and aggregate metrics only; its publication GateCheck always fails. The synthetic unit test exercises composition and privacy without presenting mock results as model evidence.

The model-promotion gate remains closed until artifacts, license, provenance, independent evaluation and the immutable public revision are verified. This does not block the library/SDK release. Adding a models.jsonl default alias before those checks would make an unvalidated checkpoint appear shipped.

Paid compute and public model publication require a separate owner decision. The cost plan must specify hardware, provider rate, estimated training and export hours, storage and egress, a spending ceiling, and evidence from a small local pilot. No paid work is configured or started here.