Synthetic section-label training-data builder for the DocType/Section family #1826

Summary

Epic OM-052 needs assembled training data for the DocType/Section model family before any training run, but the only synthetic generator today is training/synthetic/locale_phi.py for PHI. This sub-task adds a synthetic section-label corpus builder that emits token-classification examples (BIO section-boundary tags + a note-type label) from templated clinical notes, producing a leakage-checked JSONL dataset the recipe can consume. Actual GPU training is out of scope.

Scope

• Add openmed/training/synthetic/section_labels.py generating templated notes with gold section boundaries (from clinical/sections.py canonical labels) and note-type labels.
• Emit BIO-style section-boundary token tags plus a document-type label per example as JSONL, matching the corpus.py record contract.
• Support a reproducible seed so a fixed seed yields byte-identical output; expose a build_section_dataset(seed, n) entrypoint.
• Run a leakage check that no example text overlaps the synthetic eval fixtures (leakage-first) and stamp a dataset manifest.
• Ensure only synthetic/public content is generated; DUA corpora are eval-only and never written to disk here.
• Add unit tests for determinism, label validity against canonical section labels, and the leakage assertion.

Acceptance criteria

• build_section_dataset(seed, n) writes JSONL with per-token section BIO tags and a document-type label per example.
• A fixed seed produces byte-identical output across two runs (determinism unit-tested).
• Every emitted section label is a member of the canonical section label set; invalid labels raise.
• A leakage check asserts zero text overlap between generated data and the synthetic section eval fixtures.
• New unit tests pass with .venv/bin/python -m pytest tests/unit/training/ -q.
• No DUA or non-synthetic text is written; the builder is offline and deterministic.

Out of scope

• R
• u
• n
• n
• i
• n
• g
• G
• P
• U
• t
• r
• a
• i
• n
• i
• n
• g
• ;
• t
• h
• e
• t
• r
• a
• i
• n
• i
• n
• g
• c
• o
• n
• f
• i
• g
• /
• h
• a
• r
• n
• e
• s
• s
• (
• s
• e
• p
• a
• r
• a
• t
• e
• s
• u
• b
• t
• a
• s
• k
• )
• ;
• t
• h
• e
• a
• c
• t
• u
• a
• l
• D
• o
• c
• T
• y
• p
• e
• /
• S
• e
• c
• t
• i
• o
• n
• c
• h
• e
• c
• k
• p
• o
• i
• n
• t
• a
• n
• d
• i
• t
• s
• g
• a
• t
• e
• s
• ;
• R
• e
• l
• E
• x
• /
• L
• i
• n
• k
• f
• a
• m
• i
• l
• i
• e
• s
• .

Files

• openmed/training/synthetic/section_labels.py
• openmed/training/corpus.py
• tests/unit/training/test_synthetic_section_labels.py

Task: OM-818 · Milestone: v2.2 · Priority: P1 · Size: M
Depends on: OM-052 · Blocks: OM-819
Roadmap: v2.2 trustworthy clinical data exchange follow-on wave
Spec: PLANS/V2/EXECUTION/tasks/OM-818.md