Lucius-Morningstar/mailroom-dataset

Mailroom Dataset v1

Lineage: this is the standalone successor of Lucius-Morningstar/mailroom-corpus (frozen v8 baseline, 2,000 rows — never destroyed, plan §4; hardened release pinned at revision eafe1ab4c0d330d8f9c7a5fb254155e75d290828). It is canonically referred to as v9 of the mailroom corpus family; within this repository's own lineage it is versioned v1 (dataset_version, plan §62). The original corpus remains available and pinned; evaluation traces that target the successor should record this repo + the parent revision above.

What this is (and is not)

mailroom-dataset is an evaluation/ingress simulation corpus for LLM-Mailroom: it simulates documents arriving at a mailroom pipeline so that classification, routing, extraction, grouping, adjudication, and recovery can be measured end-to-end. It is not a model-training dataset, and its component corpora are deliberately heterogeneous (plan §1, §78).

Splits: the md5(filename) 90/10 split is an evaluation partition for reproducible sampling — NOT ML train/test semantics. The corpus is not organized around training; the primary unit is the evaluation scenario (plan §48–49).

Label separation (LLM-safe formatting): the default (blind) config carries exactly 4 columns — filename, doc_text, prompt, metadata — and never contains a label column. All ground truth lives in the ground_truth config, joined on filename. An LLM processing the blind config cannot see labels, intent, expected classes, or clause annotations.

v8-inherited metadata aggregates: 509 v8 contract rows carry a clause_count and 152 v8 merger rows a maud_label_count integer inside their metadata blob (inherited verbatim from mailroom-corpus v8 to preserve zero identity drift). These are label-derived aggregate counts, not labels, and are not present in any v9 expansion row. Consumers that require strict label-free metadata may treat them as weak indirect signals; they cannot be stripped without violating the zero-drift mandate.

GT completeness: every expected ground-truth field is populated for all 3,302 rows — label_evidence is derived from the CUAD/MAUD clause annotations on the 509 contract / 152 merger rows; list-typed fields (denial_reasons, supporting_documents, relationships, related_document_ids) carry a valid JSON array ([] = no items); clause labels are JSON objects ({} = no annotations on the 91 SEC EDGAR EX-10 contracts). The single documented allowance is adjuster: the CMS DE-SynPUF, GNOTHEIA property, and INSURBIAS subclasses have no adjuster in their sources, so those 950 rows leave adjuster empty (BDR auto rows carry their pseudonyms). Three outpatient :2 notices with service dates literally "N/A" in the source carry the verbatim N/A marker. The 150 INSURBIAS auto rows (issue #29) ship claim narratives only, so their supporting_documents are derived deterministically from each narrative's referenced features — repair estimate on any vehicle-damage / repair assertion (the v8 BDR auto precedent), plus police report / damage photos / medical records / fire report on explicit feature matches; the 6 rows whose narrative references no supporting-document feature carry [] as a documented absence (see v9_build._insurbias_supporting_doc_absent).

Composition (v1 = v8 + expansions)

Document class	Rows	Share	Δ vs v8
contract	600	18.2%	+91
corporate_record	450	13.6%	+411
correspondence	1,000	30.3%	+650
insurance_claim	1,100	33.3%	+150
merger_agreement	152	4.6%	+0

Expansion draws (deterministic, sha256-within-stratum):

| contract | +91 | | corporate_record | +411 | | correspondence | +650 | | insurance_claim | +150 |

Configs

Config	Rows	Contents
default (blind)	3,302	filename, doc_text, prompt, metadata — zero labels by construction
ground_truth	3,302	labels (27-key schema), identity/hashes, evaluation contract, matter/group
bundles	50	§14 synthetic bundle families over real anchors (flagged synthetic_constructed)
streams	62	§27–29 interleaved ingress stream (RUN-SIM-001) with distractors
fixtures	32	§68–§72A recovery/adversarial fixtures (calibration quartet, arbiter, failure stages)

Sources & licensing

Class	Source	License
contract	CUAD v1 (509) + SEC EDGAR EX-10 (91)	CC BY 4.0 / US public domain
merger_agreement	MAUD v1 (Zenodo 7500064)	CC BY 4.0
corporate_record	SEC EDGAR S-1/8-K exhibits	US public domain
correspondence	Enron deduplicated corpus (247,523-row pool)	research-use (other); real named-individual PII — conservative handling
insurance_claim	CMS DE-SynPUF + GNOTHEIA + BDR + INSURBIAS narratives	Apache-2.0 / MIT / CC BY 4.0

Usage:

from datasets import load_dataset
blind = load_dataset("Lucius-Morningstar/mailroom-dataset", "default")
gt = load_dataset("Lucius-Morningstar/mailroom-dataset", "ground_truth")
# join on "filename" to pair documents with their labels