Lucius-Morningstar/mailroom-dataset

Mailroom Dataset

3,302 real and synthetic business documents across five classes (contracts, merger agreements, correspondence, insurance claims and corporate records) for evaluating document triage: classification, routing, field extraction, grouping and recovery. Documents arrive blind; ground truth is kept in a separate config so a model under test never sees a label.

Current release: v9.3 (Hub tag v9.3). Pin a tag in evaluations; main moves.

class	train	test	total	subclasses
contract	540	60	600	26
corporate_record	403	47	450	10
correspondence	915	85	1,000	8
insurance_claim	986	114	1,100	6
merger_agreement	135	17	152	5
all	2,979	323	3,302	55

Quick start

import json
from datasets import load_dataset

repo = "Lucius-Morningstar/mailroom-dataset"
docs = load_dataset(repo, "default", revision="v9.3")["test"]       # documents
gold = load_dataset(repo, "ground_truth", revision="v9.3")["test"]  # labels, identity, eval contract
gold_by_name = {r["filename"]: r for r in gold}

# The model sees text and prompt only: filename and metadata can reveal answers.
# Keep each filename beside its prediction, outside the model input, for the join.
for name, doc in zip(docs["filename"], docs.select_columns(["doc_text", "prompt"])):
    prediction = my_model(doc["doc_text"], doc["prompt"])          # your model here
    row = gold_by_name[name]
    labels = json.loads(row["gt_fields"])                           # gt_fields is a JSON string
    presence = json.loads(labels["gt_presence"])                    # per-field presence code
    # score prediction against row["expected"], labels and presence

Configs

config	rows	contents
default	3,302	filename, doc_text, prompt, metadata. No label columns, but see the metadata caveat under Bias and limitations.
ground_truth	3,302	class and subclass, gt_fields (JSON), identity and hashes, the evaluation contract, matter grouping and provenance
bundles	50	synthetic multi-document families built over real anchor documents (flagged synthetic_constructed)
streams	62	one interleaved ingress stream (RUN-SIM-001) over the bundle matters, with 12 unrelated distractors
fixtures	32	recovery and adversarial cases: calibration quartet, arbiter scenarios, failure stages

ground_truth columns (35)

group	columns
labels	expected, expected_subclass, gt_fields
identity	document_id, content_sha256, normalized_text_sha256, source_corpus, source_document_id, source_filename, source_revision
evaluation contract	expected_specialist, expected_stage, review_expected, review_reason, retry_expected, expected_post_retry_state
provenance	annotation_source, annotation_method, annotation_model, annotation_prompt_version, annotation_confidence, annotation_reviewer, annotation_timestamp
matter and grouping	matter_id, matter_construction, group_id, group_role, thread_position, thread_size, thread_evidence, relationships, related_document_ids
other	filename, prompt, split

gt_fields holds the class's label fields plus gt_presence, token_estimate, context_window_band (<=4k, <=16k, <=32k, >32k), clause_count and maud_label_count. CUAD and MAUD clause annotations are in cuad_clause_labels and maud_clause_labels.

On the 350 GNOTHEIA and BDR insurance rows, six fields hold the string "nan" as an absence placeholder, not a value: content_topic, label_evidence, sentiment_evidence, sentiment_label, sentiment_score and topic_evidence (2,100 cells). Treat them as absent, and use gt_presence where it has a code. The "N/A" dates on 3 CMS outpatient claims are the source's own marker; keep them.

Presence codes

Every label-bearing field in gt_presence carries one code, so a blank value is never ambiguous:

code	meaning	cells
populated	a real expected value	23,453
schema_documented_absence	the field applies, but this document has none (e.g. denial_reasons on an approved claim)	2,020
pending_annotation	not labeled yet (the 91 SEC EDGAR EX-10 contracts' CUAD clauses)	91
not_applicable	the class never carries this field	41,758

A field missing from gt_presence is unlabeled. Scorers should exclude it from both sides, so neither a filled nor an empty prediction counts. On corporate records, the per-subclass entity fields from llm-dojo-scoring's corporate-record schema stay unlabeled until they are annotated.

Splits

A row is in test when md5(filename) % 10 == 0 (≈10%). The rule is stable across releases. The splits are an evaluation partition for reproducible sampling, not a training/validation design: this is not a training corpus.

Sources and licenses

class	source	rows	license
contract	CUAD v1 (theatticusproject/cuad)	509	CC BY 4.0
contract	SEC EDGAR EX-10 exhibits	91	public SEC filings; company-authored, not US-government works (see below)
merger_agreement	MAUD v1 (Zenodo 7500064)	152	CC BY 4.0
corporate_record	SEC EDGAR S-1 / 8-K exhibits	450	public SEC filings; company-authored, not US-government works (see below)
correspondence	Enron, deduplicated (enron-correspondence-dedup)	1,000	research use; contains real names
insurance_claim	CMS DE-SynPUF (synthetic Medicare claims)	600	US public domain
insurance_claim	GNOTHEIA synthetic property FNOL	200	Apache-2.0
insurance_claim	BDR motor-claim decisions	150	MIT
insurance_claim	INSURBIAS claim narratives	150	CC BY 4.0

The licensing is mixed, so the Hub metadata says mixed-per-source. CC BY 4.0 covers this compilation and its labels; each document keeps its source's terms, as listed above. SEC EDGAR filings are free to access and reuse under the SEC's public-access terms, but the companies wrote them, so they are not US-government works and this card does not claim they are in the public domain. The Enron emails are released for research use and contain real names. CC BY-NC sources were excluded.

Annotation

Labels come from six regimes, recorded per row in annotation_method:

annotation_method	rows	what it means
synthetic	1,100	generated with the document (insurance sources)
source_native	661	the source corpus's own expert labels (CUAD, MAUD)
heuristic	646	deterministic rules over the document or its metadata
llm_zero_shot	637	deepseek-chat zero-shot (correspondence intent), with confidence and status
verified_join	162	exact sha256 body match to AESLC/Enron gold
human_annotated	96	a person's label

Correspondence intent uses a closed 8-class vocabulary: payment_demand, notice, analysis, request, update, meeting_invite, press_communication, other. All eight occur in the test split.

Personal information

The correspondence class is real Enron email: it contains the names and addresses of real people. Use it for evaluation only, and do not use it to profile individuals. The insurance claims are synthetic or anonymized; contracts, mergers and corporate records are public filings.

Bias and limitations

• Contracts, mergers and corporate records are US public-company filings; other jurisdictions are rare.
• The insurance claims are synthetic and simpler than real claim files.
• 742 of the 1,000 correspondence intents are machine labels (637 llm_zero_shot) or heuristics (105); 258 are human or exact gold joins. intent_source and intent_status say which, and 25 rows are flagged_review.
• On 411 SEC EDGAR corporate records, keywords holds only the record type and the entity name. The v9.2.1 repair removed the exhibit, form, accession and CIK strings that filled it before. Treat these keywords as not yet labeled.
• Corporate-record entity_name is the EDGAR filer name, not always the string printed in the document (about a quarter differ, mostly Co. vs Company).
• Blind metadata is not fully blind. Some source fields in default's metadata give away the answer without reading the text: the subclass (category, lob, claim_subtype, exhibit_description), the class (which keys are non-empty, and source-style filenames such as carrier:…), insurance claim_number (source_row_id), and on the 152 merger agreements the per-category MAUD annotation counts (maud_categories). To score a model fairly, give it doc_text and prompt only. v10 removes metadata and filenames from the blind config.

Versions

tag	change
v9.3	Corporate records gain per-subclass presence keys; merger agreements route to merger_agreement_specialist (labels only; rows and text unchanged). Labels are versioned separately (labels_version in manifest.txt).
v9.2.1 (branch)	Corporate-record label repair: entity_name, record_type, grounded subject_matter
v9.2	Six-regime annotation provenance; related_document_ids populated; build-internal column removed
v9.1	Label-derived counts moved out of the blind config; presence codes; context-window bands
v9	3,302 rows: correspondence, corporate records, EX-10 contracts and insurance narratives added to the v8 base
v8	mailroom-corpus, 2,000 rows (frozen at eafe1ab4)

Every release since v9 keeps the row set, document_ids and document text unchanged. Release notes: CHANGELOG.md. Byte checksums for every file are in manifest.txt.

Citation

@misc{mailroom_dataset,
  title        = {Mailroom Dataset: An Evaluation and Ingress Simulation Corpus for Legal Document Triage},
  author       = {Morningstar, Lucius},
  year         = {2026},
  howpublished = {Hugging Face: \url{https://huggingface.co/datasets/Lucius-Morningstar/mailroom-dataset}}
}

Please also cite the source corpora you use: CUAD (arXiv:2103.06268) and MAUD (arXiv:2301.00876).

Issues and pull requests: Exios66/Mailroom-Corpus-EDA.