mailroom-modernbert-training

Cleaned + prepared hierarchical-classification training set for the ModernBERT-base ingest fast-path — the fine-tuning surface of the mailroom-ml synthetic-data layer.

The layer (end to end)

Lucius-Morningstar/mailroom-dataset      corpus (GT labels, canonical v9)
        │  (working copy, pinned)
        ▼
THIS REPO (mailroom-modernbert-training) @ pinned revision
        │  training/train_modernbert.py (--data <repo>)
        ▼
Lucius-Morningstar/mailroom-modernbert-classifier   (trained checkpoint)

Synthetic enrichment (tier 1/2/3) is assembled by training/assemble_enrichment.py from external pools (enron, cms, gnotheia, bdr, insurbias, blind) with per-tier caps, a 7-gate audit, and example_weight/lineage/tier` columns. No enrichment tier is adopted into this dataset yet — adoption is gated by the plan's §6.1 A/B ladder (an adopted tier must beat the no-enrichment baseline on the held-out test split before it ships here). This revision is the pure-canonical baseline.

Rows

Documents	3302 (train 2680 / validation 299 / test 323)
Windows (8,192-token)	4989 (train 4499 / validation 490)
doc_type classes	5 (+ unknown, inference-only)
Subclass heads	5 per-class heads
Model target	answerdotai/ModernBERT-base (max_tokens=8192, overlap=512)
Source revision	19720ceb4e29bc3134a88507aa57cdfac7a64a1b (GT-closure)

Configs

• documents — one row per document: filename, document_id, content_sha256, source_revision, title, doc_text, doc_type, subclass, corpus_split, split (train/validation/test), token_estimate. The held-out test split (323 rows) never touches training — it is the eval surface.
• windows — one row per pre-computed 8,192-token ModernBERT window (title + "\n\n" + window, 512-token overlap) for train + validation only — the fine-tune surface. Test stays document-level.

Labels

• doc_type — 5 classes: contract, merger_agreement, corporate_record, correspondence, insurance_claim.
• subclass — canonical key per class, normalized through the same mapping the sandbox corpus uses (llm_dojo_scoring.corpus.normalize_corpus_subclass, DMR-066; vendored in mailroom_ml/labels.py): Service/service unify, Co_Branding → co_branding, Joint Venture _ Filing → joint_venture. Per-class vocabularies in vocabularies.json; per-head id2label + class weights (inverse-frequency over the train split) in labels.json.

Splits

• Corpus train (2,979) → 90/10 stratified train/validation (stratified by doc_type, seed 42, deterministic RandomState shuffle — no sklearn dependency, byte-identical rebuilds).
• Corpus test (323) → held out entirely.

Input construction

title follows the corpus's title-wins convention: subject → exhibit_description → empty (semantic-only). The filename fallback was removed on 2026-09-20) and a real pipeline filename does not encode the class. Each window input is title + "\n\n" + window_text (body-only when there is no semantic title), truncated to 8,192 tokens.

Text is run through the pipeline's deterministic intake clerk (deterministic_normalize, vendored in mailroom_ml/normalize.py) at build time — the same clerk llm-mailroom apply_intake runs before the classifier — so training input is byte-representative of inference input (NFC, newline unify, NBSP, zero-width strip, C0 controls, hyphen unwrap, blank-run collapse, horizontal collapse, edge trim). A filename_leak_audit gate in verify_stage fails the build if a filename-derived title ever returns.

Provenance

Built by mailroom_ml (src/) + training/build_dataset.py in the mailroom-ml repo. manifest.txt carries the build facts + sha256s; rebuilds are byte-identical (sorted rows, seeded split, pinned tokenizer, no timestamps in artifacts). Publishing is operator-only (--publish) and byte-verifies every sidecar against the Hub after upload.