Mailroom-Corpus-EDA: appendix

Published v9.2 tables · configuration, provenance, sources and licences

These tables support the Mailroom-Corpus-EDA case study. They describe the published v9.2 revision of mailroom-dataset⁠ and come from its dataset card. No documents or participant-level records are reproduced here.

Published configurations

Configuration	Rows	Contents
default	3,302	Four columns only: filename, document text, prompt, metadata. No label column.
ground_truth	3,302	Labels, identity hashes, evaluation contract and matter grouping, joined on filename.
bundles	50	Synthetic bundle families over real anchors, flagged as constructed.
streams	62	Interleaved ingress stream with distractors.
fixtures	32	Recovery and adversarial fixtures.

ground_truth carries 35 top-level columns; the nested gt_fields object holds 34 keys, 32 scalar and two matter lists.

Class composition

Document class	Rows	Share
Insurance claim	1,100	33.3%
Correspondence	1,000	30.3%
Contract	600	18.2%
Corporate record	450	13.6%
Merger agreement	152	4.6%

Sources and licences

Class	Source	Licence
Contract	CUAD v1 (509) plus SEC EDGAR EX-10 exhibits (91)	CC BY 4.0; US public domain
Merger agreement	MAUD v1, Zenodo 7500064	CC BY 4.0
Corporate record	SEC EDGAR S-1 and 8-K exhibits	US public domain
Correspondence	Deduplicated Enron corpus, 247,523-row pool	Research use; contains real named-individual PII
Insurance claim	CMS DE-SynPUF, GNOTHEIA, BDR and INSURBIAS narratives	Apache-2.0, MIT, CC BY 4.0

These are the component terms. Citing the combined corpus does not replace them; see Cite the dataset.

How labels were produced

Annotation method	Rows
Synthetic	1,100
Source-native	661
Heuristic	646
LLM zero-shot	637
Verified join	162
Human annotated	96

Every label-bearing field also carries a presence code, so an empty value is never ambiguous:

Presence code	Fields	Share
Populated	22,591	34.2%
Not applicable	41,338	62.6%
Schema-documented absence	2,020	3.1%
Pending annotation	91	0.1%

No field is an undocumented gap. The 91 pending entries are the SEC EDGAR EX-10 contracts, whose CUAD clause labels wait on a separate annotation pass.

Context-window bands

Band	Rows
Up to 4k tokens	2,530
Up to 16k	428
Up to 32k	120
Over 32k	224

Split and held-out test

The reproducible split puts a document in the test set when the MD5 of its filename, modulo 10, is zero. That gives 323 test documents and 2,979 for training and validation, split 2,680 / 299 at 90/10 and stratified by document type with seed 42. The split is an evaluation partition, not a claim that the corpus was designed as a training set.

Scope