Lucius-Morningstar/mailroom-corpus

Successor: the v9 build of this corpus family is published as Lucius-Morningstar/mailroom-dataset (v1, 3,302 rows: insurance_claim 1,100 · correspondence 1,000 · contract 600 · corporate_record 450 · merger_agreement 152; 5 configs incl. blind default). This repo remains the frozen v8 baseline (2,000 rows, revision eafe1ab4c0d330d8f9c7a5fb254155e75d290828) — never destroyed.

Docclass Merged Corpus

Single flat document-classification surface: 2,000 legal documents across five corpora, one row per document (schema v6):

Corpus	Rows	doc_type	Source
CUAD contracts	509	contract	CUAD v1 (CC BY 4.0, The Atticus Project)
MAUD merger agreements	152	merger_agreement	MAUD v1 (CC BY 4.0, Wang et al. 2023)
S-1 corporate-record exhibits	39	corporate_record	SEC EDGAR public filings
Enron correspondence sample	110	correspondence	Lucius-Morningstar/enron-correspondence-dedup (CMU Enron Email Dataset, research-use)
CMS DE-SynPUF rendered EOBs	750	insurance_claim	Lucius-Morningstar/cms-desynpuf-insurance-claims (CMS DE-SynPUF Sample 1 via Exios66/claims-data-eda)

Source datasets & original references

Four public corpora, one flat surface. Every row self-identifies its origin via metadata.source / metadata.source_dataset.

CUAD contracts — 509 rows (contract)

• Corpus: CUAD v1 — the Contract Understanding Atticus Dataset, The Atticus Project, Inc., released CC BY 4.0.
• Original download: https://www.atticusprojectai.org/cuad · mirror repo: https://github.com/TheAtticusProject/cuad
• Reference: Hendrycks, Burns, Chen & Ball, "CUAD: An Expert-Annotated NLP Dataset for Legal Contract Review" (NeurIPS Datasets & Benchmarks 2021, arXiv:2103.06268).
• Ingestion: byte-verified export of the Braintrust mirror (mailroom-cuad-contracts-full). Rows carry CUAD's own 28-group contract taxonomy as expected_subclass (e.g. Co_Branding, Distributor) plus clause counts and applicable-categories metadata.

MAUD merger agreements — 152 rows (merger_agreement)

• Corpus: MAUD v1 — the Merger Agreement Understanding Dataset, The Atticus Project, Inc., released CC BY 4.0 (152 agreements, 47,000+ expert labels over the ABA 2021 Public Target Deal Points Study).
• Original download: https://www.atticusprojectai.org/maud/
• Reference: Wang, Scardigli, Tang, Chen, Levkin, Chen, Ball, Woodside et al., "MAUD: An Expert-Annotated Legal NLP Dataset for Merger Agreement Understanding" (EMNLP 2023, arXiv:2301.00876).
• Ingestion: agreement texts streamed from the Zenodo v1 corpus export; consideration-type annotations inform expected_subclass, label counts carried in metadata.

S-1 corporate-record exhibits — 39 rows (corporate_record)

• Source: exhibits extracted live from SEC EDGAR public filings — https://www.sec.gov/edgar.shtml.
• Public regulatory filings (US government works); exhibit type/description and filer/accession metadata retained per row.

Enron correspondence sample — 110 rows (correspondence)

• Source corpus: the CMU Enron Email Dataset — Bryan Klimt & Yiming Yang, Carnegie Mellon University, 2004. Original download: https://www.cs.cmu.edu/~enron/
• Reference: Klimt & Yang, "The Enron Corpus: A New Dataset for Email Classification Research" (ECML 2004, pp. 217–226).
• Ingestion: drawn deterministically (sha256(filename) order) from the sha256-verified parquet shards of Lucius-Morningstar/enron-correspondence-dedup, the family's deduplicated/enriched Enron derivative. Per-email provenance (custodian, folder, date, sender, message-id) rides along in metadata; research-use terms inherited — see License notes below.

Schema v5 additions (KANBAN-084, 2026-08-23)

• New class: insurance_claim (+400 rows, subtypes inpatient / outpatient / carrier / pde) from cms-desynpuf-insurance-claims — rendered EOBs with a verbatim GT contract aligned to llm-mailroom's InsuranceClaimExtraction. Synthetic data; PAID claims only (coverage_determination always "approved", denial_reasons always empty); adjuster always null; single line of business (health). CMS caveat: evaluation substrate, not epidemiology.
• Split reconciliation: the claims source keyed its own file placement on md5(record_id); this dataset applies the FAMILY rule (md5(filename) % 10 == 0 → test) to every row — 65/400 claims rows changed placement vs the source repo. The split column here is authoritative.
• Clause-level GT (ground_truth config ONLY — never in the blind config):
  • cuad_clause_labels on contract rows: the official CUAD annotation set (machine-readable superset of masterlabels.csv) — 509/509 contracts joined; 13,753/13,753 answer spans verified at exact char offsets against the stored doc_text. Compact JSON: clause name → [{text, start}].
  • maud_clause_labels on merger_agreement rows: MAUD gold labels across its classification tasks (category / answer / valid_classes / label_idx) — 152/152 contracts joined via contract id.
  • These are the scoring substrate for entity-extraction evaluation; treat them exactly like the other GT columns (separation of concerns, not encryption).

Schema v7 additions (issue #5, 2026-08-31)

• Correspondence intent hydration: every correspondence row now carries a non-null intent from the closed 8-class vocabulary (payment_demand, notice, analysis, request, update, meeting_invite, press_communication, other) — 100.0% coverage across 350 rows, plus three provenance columns on the ground_truth config: intent_source (manual | aeslc_join | llm_zero_shot), intent_confidence (0..1), intent_status (manual | auto_labeled | flagged_review).
• Hydration provenance: intent_source records the hydration PATH (disjoint values summing to 350): 0 rows manual (purpose-GT push), 0 rows aeslc_join — hydrated through the sha256 exact-body join against the Enron mail corpus (snoop2head/enron_aeslc_emails, 535k mails) and AESLC (Yale-LILY/aeslc); the mirrors carry NO intent annotations (verified 2026-08-31), so the join supplies provenance + the recovered subject_line used as constrained context — and 0 rows llm_zero_shot (constrained zero-shot LLM pass, OpenRouter deepseek/deepseek-chat, temperature 0.1, closed vocabulary).
• Confidence thresholding: confidence >= 0.85 -> auto_labeled; below -> flagged_review (0 rows flagged for the manual review queue). Non-conforming / residual rows fall to other (0 rows) — never null.

Schema v8 additions (HUB-028, 2026-09-02)

• Insurance LOB expansion: +200 property rows (GNOTHEIA synthetic polycontexts, Apache-2.0 — FNOL documents stratified by loss event: fire/water/storm/burglary/…) and +150 auto rows (BDR motor-claims decisions, MIT — decision letters stratified by accident type × APPROVE/REVIEW/REJECT, all reject rows included). Both carry the full 27-key GT: claim ref, policy, insured party, dates, claimed amount, adjuster pseudonym (auto), damages narrative, coverage determination (property = pending, honest — no adjudication in the source; auto = approved/pending/denied), feature-grounded denial reasons (auto reject), intent, subject_matter, keywords + provenance.
• Full GT conformance: all 1300 rows now carry intent / subject_matter / keywords / intent provenance — including all 600 CMS DE-SynPUF rows (was 246/600 missing subject/keywords, 600/600 missing intent).
• Metadata for all entries: source_dataset / source_revision / source_row_id / lob / peril / license ride every row; union-normalized cast-safe metadata.
• Test-split nullification: every test-split row carries fully populated applicable GT (no None/NaN; '' allowed only for schema-documented absence, e.g. adjuster on CMS/property rows).

Mailroom evaluation hardening (v0.2/v0.3/v0.4, 2026-09-02)

One revision delivering three §84 markers (plan §26: conceptual targets, not forced releases). Blind default config is UNCHANGED (4 columns, no labels).

v0.2-mailroom-hardened — new columns on ground_truth:

• Identity/hashes (§9–§12): document_id (stable DOC- + sha256[:16] over source corpus + source filename — never row order, never content), source_corpus, source_document_id, source_filename, source_revision ('' until tracked by a builder), content_sha256 (canonical doc_text bytes), normalized_text_sha256 (NFC, LF-folded, whitespace-collapsed — the duplicate-detection key).
• Evaluation contract (§31/§43/§57–59): expected_specialist (live taxonomy.yaml registry), expected_stage (archived for canonical rows), review_expected/review_reason (false/'' on canonical rows — closed §73 vocabulary reserved for fixtures), retry_expected/ expected_post_retry_state, annotation_source/annotation_method (verified_join 162 / llm_zero_shot 92 / human_annotated 96 / synthetic 950 (DE-SynPUF 600 + GNOTHEIA 200 + BDR 150) / source_native 700) /annotation_model/ annotation_prompt_version/annotation_confidence/annotation_reviewer/ annotation_timestamp. Built on the v8 base (2,000 rows, HUB-028): the v8 LOB rows carry their own dataset as source_corpus / annotation_source (GNOTHEIA / BDR) and their pinned upstream source_revision; published v7 document_ids are unchanged (0 drift).

v0.3-matter-aware — §14A methodology (verified 2026-09-02 against the raw CMU maildir: In-Reply-To/References are structurally ABSENT — 0/350 raw files, 0/247,523 upstream dedup rows — so header-thread ground truth cannot exist in this corpus family):

• ground_truth gains matter_id, matter_construction, group_id, group_role, relationships (list), related_document_ids (list), thread_position, thread_size, thread_evidence. Populated ONLY via heuristic_reconstructed (normalized subject + custodian + 30-day window, degenerate subjects excluded): 19 rows in 7 threads; 1,981 rows unassigned (all threads are correspondence; the v8 insurance LOB rows carry no header-thread evidence) — honest baseline, counted separately, never merged into a "matters" total. source_native_thread stays implemented (guarded) for future feeds that carry real reply headers.
• NEW bundles config (§14 synthetic families, flagged synthetic_constructed): five family templates over REAL anchor rows (§14's legal/insurance worked examples span ≥2 classes) — legal_contract_family, insurance_claim_family, merger_family, corporate_record_family, correspondence_thread_family; seed 42, 2 anchors per family, one template_variant duplicate per instance (§87). Manufactured siblings carry synthetic: true, an explicit scaffold doc_text header, and MATTER-SYN-* filenames (no snapshot collisions); they claim NO source provenance (source_corpus = '') and annotation_method = synthetic. Roles/relationships come ONLY from the closed GROUP_ROLES/RELATIONSHIP_TYPES vocabularies.
• NEW streams config (§27–§29/§48 STREAM eval tier): the bundle members interleaved into ONE reproducible ingress stream — RUN-SIM-001, round-robin across the bundle matters (A1 B1 A2 C1 B2 ... — never matter-contiguous, §28), with distractor rows injected every 4 positions (real corpus rows from outside every matter, matter_id/group_id empty, §29). Every row carries simulation_run_id + sequence_position (strictly increasing, §27 reproduces the exact incoming sequence).

v0.4-recovery-suite — NEW fixtures config (§68–§72A fixture content; fixture: filename namespace, so these are evaluation scenarios, not corpus rows):

• §70 calibration quartet — correct_high / correct_low / wrong_high / wrong_low for ALL FIVE classes, each row probing the LIVE routing bands from llm-mailroom's taxonomy.yaml (probe sits just inside the band edge); cell→fixture-kind mapping is closed (wrong_high is the silent-archive failure mode under test).
• §72A arbiter scenarios — one per closed outcome (stands, re_extract, escalate_human_review) + a review-correction scenario.
• §58 failure-stage matrix — one minimal failure fixture per stage (ingestion → archival), the spine for first-pass vs. recovered success.
• All review/retry expectations are DERIVED through the same mailroom_eda.eval_contract module that produces the v0.2 columns — the fixtures and the corpus share one evaluation contract.

Splits: md5(filename) % 10 == 0 → test applied uniformly (existing rows keep their published split; new synthetic rows derive theirs the same way). Scaffold modules live in the monorepo (packages/mailroom-corpus-eda/src/mailroom_eda/); methodology in docs/DOCCLASS_CONTRACT.md §9/§9A/§9B/§9C.

Original files (KANBAN-105 addendum, 2026-08-30)

The upstream originals for the three corpora that have them ride along under files/ for easy access — the text content is what agents see; these are a convenience layer, not load-bearing:

doc_type	files	form	source
contract	0	source PDFs (metadata.pdf_path layout)	theatticusproject/cuad (CC BY 4.0)
merger_agreement	0	upstream contract_N.txt (MAUD ships no PDFs)	Zenodo 7500064 (CC BY 4.0)
corporate_record	0	EDGAR exhibit originals (.htm)	SEC EDGAR via metadata.exhibit_url (public domain)

metadata.original_file carries the Hub-relative path on every row that has one (cast-safe "" elsewhere: correspondence rows are maildir text and insurance_claim rows are synthetic renders — the render IS the original). Fetch one: hf_hub_download("Lucius-Morningstar/mailroom-corpus", "files/contract/Part_I/License_Agreements/<file>.pdf", repo_type="dataset"). Per-file sha256 + sizes: original_files_mapping.jsonl sidecar.

Schema v6 additions (KANBAN-105, 2026-08-30)

• Correspondence rebalance: +0 rows (110 → 110) — deterministic sha256(filename) stratified draw from enron-correspondence-dedup after excluding every existing filename; the shared Enron labelers (subclass / content-topic / sentiment) were RE-RUN on every drawn row as a verification pass and reproduce the Hub ground truth exactly; the KANBAN-103 phrase-lexicon GT overrides are honored. The dedup corpus carries no attorney_demand rows beyond the 3 already present (all in the v4 sample) — honest gap, not an omission.
• Insurance boost: +350 rows (400 → 750) — newly rendered EOBs from CMS DE-SynPUF Sample 1 via Exios66/claims-data-eda with the verbatim GT contract asserted at render time; every existing record_id was excluded, so the original 400 claims are untouched. Subtypes: auto, carrier, inpatient, outpatient, pde, property. Same synthetic-data caveats as v5 (PAID claims only, adjuster null, health LOB).
• Blind-surface repair: the default (blind) config no longer carries the label equivalents expected_doc_type / expected_subclass inside metadata (a v4-era flat-dump artifact v5 shipped verbatim) — it now honors the card's "NO label columns" contract; labels live ONLY in the ground_truth config (0 rows repaired).
• Purpose/gist GT: the ground_truth config now carries intent / subject_matter / keywords columns on BOTH splits (train rows labeled by the llm-mailroom purpose-GT push of 2026-08-30; new append rows are empty until the incremental labeler pass fills them in a follow-up revision — then every corporate_record / correspondence / insurance_claim row is gradable against the controlled INTENT_LABELS vocabularies).
• Class balance after v6: contract 509 (25.4%), insurance_claim 950 (47.5%), correspondence 350 (17.5%), merger_agreement 152, corporate_record 39.

⚠️ Two-config layout (schema v4): labels live in ground_truth

As of v4 the dataset follows the same agent-blind pattern as enron-correspondence-dedup:

from datasets import load_dataset

# what a classification agent may see — NO label columns:
blind = load_dataset("Lucius-Morningstar/docclass-merged")

# what the scorer joins against (explicit opt-in), keyed 1:1 on filename:
gt = load_dataset("Lucius-Morningstar/docclass-merged", "ground_truth")

Both configs are served as pre-sharded parquet (parquet/<config>/<split>/*.parquet, zstd). The complete legacy surface — all 700 pre-v4 rows plus the 110 correspondence additions in the original combined row shape — also ships line-delimited as docclass_merged.jsonl for pipelines that prefer raw JSONL; it is row-for-row identical to the parquet configs joined on filename. The prior v3 build remains recoverable from git history (previous_fingerprint_v3 in manifest.txt).

Row shape

default (blind) config — one row per document:

• filename — the source FILE name (CUAD PDF basename, MAUD/S-1 dump filename, or Enron maildir path)
• doc_text — full document text
• prompt — reserved task prompt (empty string on every row, matching the v3 convention)
• split — train/test, assigned deterministically by md5(filename) mod 10 == 0 -> test; stable across rebuilds and identical to the rule used by sibling datasets
• metadata — provenance struct, all values plain strings (32-key union; per-corpus fields, empty string when not applicable). Enron rows carry custodian/folder/date/sender/message-id provenance plus their original license note and the sampling method.

ground_truth config — keyed 1:1 on filename:

• All rows: expected (gold doc_type), expected_subclass (second-level gold), split
• Legacy rows (contract/merger_agreement/corporate_record): exactly the v3 label columns — nothing else
• Correspondence rows additionally carry the enriched-evidence columns from enron-correspondence-dedup: label_evidence, content_topic, topic_evidence, sentiment_score ∈ [-1, 1], sentiment_label, sentiment_evidence

The wider column set exists only on correspondence rows; the parquet schemas declare these columns explicitly nullable across both configs (partial-null schemas crash the Hub viewer's conversion when left implicit — lesson learned in v2, see Provenance).

The correspondence sample

110 emails drawn deterministically (sha256(filename) ordering — rebuilds are byte-identical) from the deduplicated 247,523-row Enron corpus, stratified to cover every primary correspondence type and subject type:

• All 8 doc subtypes: attorney_demand ×3, demand ×16, email ×15, letter ×15, meeting_request ×15, memo ×15, notice ×16, press_release ×15
• All 11 content topics: general_business ×47, marketing_clients ×13, energy_market ×12, hr_personnel ×12, legal_contracts ×12, regulatory ×4, travel_logistics ×4, finance_earnings ×2, scheduling ×2, announcements ×1, it_systems ×1
• Bodies shorter than 80 chars were excluded from sampling; empty bodies never win a slot.

Topic/subclass labels are heuristic ground truth (deterministic lexicon and marker-taxonomy functions, human spot-checked upstream — see the honest-gaps notes in the dedup dataset card): single-topic assignment for multi-topic emails, head-window scanning, no sarcasm detection. Treat them as weak labels/routing priors.

Splits

Per-row split column: md5(filename) mod 10 == 0 -> test (~10%). The rule is deterministic and order-independent, shared across the whole Lucius-Morningstar family, so consumers recompute or extend splits without shipping separate files. Coverage after v4: train 727 / test 83. The correspondence additions follow the same filename-hash rule, so any given email keeps its split across both this dataset and the dedup corpus.

License notes

• The CUAD and MAUD portions are CC BY 4.0 (The Atticus Project; Wang et al. 2023); S-1 exhibits are public SEC EDGAR filings.
• The Enron correspondence sample inherits the stricter terms of its source: the CMU Enron Email Dataset is released for research use, and contains real personally identifying information of Enron employees. Those rows are flagged per-row in metadata.license ("Enron corpus — released for research use") and metadata.source_dataset. Treat the correspondence subset as research-only sensitive data: no redistribution of raw PII outside research contexts, no production/consumer use.
• Original corpus download source: https://www.cs.cmu.edu/~enron/ (Klimt & Yang, CMU 2004).

Related projects

• LLM Mailroom — the agentic email/document-triage system evaluated with this benchmark.
• enron-correspondence-dedup — the full 247K-row deduplicated corpus this sample was drawn from, with the same blind/GT config pattern and enrichment labelers.
• llm-entity-extraction — build pipelines (scripts/datasets/build_docclass_merged.py).
• Enron-Evaluation-Environment — correspondence labelers (correspondence_subclasses.py, content_topics.py, sentiment_scorer.py) and dedup rule (dedupe.py).
• Sibling datasets: legalbench-full, mailroom-cuad-contracts, mailroom-cuad-contracts-full.

Citation

If you use this dataset, please cite it along with the underlying corpora:

@misc{morningstar2026docclassmerged,
  title        = {Docclass Merged Corpus (Contracts + Merger Agreements + Corporate Records + Correspondence)},
  author       = {Lucius-Morningstar},
  year         = {2026},
  month        = {August},
  howpublished = {\url{https://huggingface.co/datasets/Lucius-Morningstar/docclass-merged}},
}

@inproceedings{hendrycks2021cuad,
  title     = {{CUAD}: An Expert-Annotated {NLP} Dataset for Legal Contract Review},
  author    = {Hendrycks, Dan and Burns, Collin and Chen, Anya and Ball, Spencer},
  booktitle = {NeurIPS Datasets and Benchmarks Track},
  year      = {2021}
}

@inproceedings{wang2023maud,
  title     = {{MAUD}: An Expert-Annotated Legal {NLP} Dataset for Merger Agreement Understanding},
  author    = {Wang, Steven H. and Scardigli, Antoine and Tang, Leonard and Chen, Wei and Levkin, Dimitry and Chen, Anya and Ball, Spencer and Woodside, Thomas and others},
  booktitle = {Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing (EMNLP)},
  year      = {2023}
}

@inproceedings{klimt2004enron,
  title     = {The Enron Corpus: A New Dataset for Email Classification Research},
  author    = {Klimt, Bryan and Yang, Yiming},
  booktitle = {European Conference on Machine Learning (ECML 2004)},
  pages     = {217--226},
  year      = {2004}
}

Provenance

v1–v3 built by llm-entity-extraction scripts/datasets/build_docclass_merged.py (KANBAN-071/073/074, 2026-08-23T17:00:04+00:00). v3→v4 fusion (2026-08-23): appended a deterministic stratified 110-row correspondence sample drawn from the sha256-verified parquet shards of enron-correspondence-dedup (LFS-verified at pull time; shard sha256s recorded in manifest.txt), restructured into the two-config blind/GT parquet layout, and preserved the legacy combined JSONL unchanged. Schema lessons carried forward: all-label columns non-null on legacy rows since v2/v3 (partial-null schemas crash the Hub viewer), explicit nullable unions for v4's widened ground-truth schema.

Maintenance

Issues and fixes: llm-entity-extraction issues or contact @Lucius-Morningstar on the Hub.