Lucius-Morningstar/docclass-pilot

Docclass Pilot Sample

Deterministic stratified pilot slice of Lucius-Morningstar/docclass-merged v5 (KANBAN-084, built 2026-08-24): every doc type and every subclass stratum present in the parent contributes at least one row; each stratum contributes at most 3 rows, selected by ascending sha256(filename) within the stratum (content-addressed order — decorrelated from the md5-keyed family split rule; rebuilds are byte-identical).

Coverage

doc_type	rows	subclasses covered
contract	78	26 (Affiliate_Agreements, Agency Agreements, Co_Branding, Collaboration, Consulting Agreements, Development, Distributor, Endorsement, Franchise, Hosting, IP, Joint Venture, Joint Venture _ Filing, License_Agreements, Maintenance, Manufacturing, Marketing, Non_Compete_Non_Solicit, Outsourcing, Promotion, Reseller, Service, Sponsorship, Strategic Alliance, Supply, Transportation)
corporate_record	11	5 (articles_of_incorporation, bylaws, other, powers_of_attorney, rights_instrument)
correspondence	24	8 (attorney_demand, demand, email, letter, meeting_request, memo, notice, press_release)
insurance_claim	12	4 (carrier, inpatient, outpatient, pde)
merger_agreement	13	5 (all_cash, all_stock, mixed_cash_stock, mixed_cash_stock_election, other)

Total: 138 rows across 48 subclass strata / 5 doc types (100% of strata present in the parent).

Configs (KANBAN-079 separation-of-concerns doctrine)

• default (blind): filename, doc_text, prompt, metadata. What the Mailroom pipeline agents see. ZERO answer-key columns.
• ground_truth: filename + expected (doc_type) + expected_subclass
  • promoted GT fields — the six legacy enrichment keys plus the 13 InsuranceClaimExtraction keys for the insurance_claim rows (null on other classes). Join 1:1 on filename; scorers use standard split joins.

from datasets import load_dataset
blind = load_dataset("Lucius-Morningstar/docclass-pilot", "default")
keys  = load_dataset("Lucius-Morningstar/docclass-pilot", "ground_truth")

This is separation of concerns, NOT encryption — the Hub is public; anyone may load the ground_truth config explicitly.

Provenance

• Parent: Lucius-Morningstar/docclass-merged @ revision 407bf55c993a006689aa1b6e03e57c428c852e98 (v5: adds the insurance_claim class).
• Claims source: Lucius-Morningstar/cms-desynpuf-insurance-claims (CMS DE-SynPUF 2008-2010 Sample 1 rendered EOBs, produced by Exios66/claims-data-eda, verbatim-GT contract aligned to llm-mailroom's InsuranceClaimExtraction).
• Splits: inherited from the family single-source rule (md5(filename) % 10 == 0 -> test). Claims-side note: the claims repo had keyed its own placement on record_id; at v5 fusion the FAMILY rule won, moving 65/400 rows — the pilot inherits the reconciled family splits.
• Builder: scripts/datasets/build_docclass_pilot.py in Exios66/llm-entity-extraction.

Honest gaps

• Heuristic ground truth on the legacy four classes (family labeler; see the parent card). Claims GT is exact-but-synthetic (verbatim renderer contract).
• SynPUF claims are PAID claims only: coverage_determination is always "approved", denial_reasons always empty, adjuster always null, single line of business (health). Fully synthetic data — evaluation substrate, not epidemiology (CMS caveat).
• Family scope is FIVE doc classes: contract, merger_agreement, correspondence, corporate_record, insurance_claim. court_opinion and compliance_filing were removed from the dataset family's scope by human directive (2026-08-23) — no upstream corpora exist. This does NOT change llm-mailroom's runtime taxonomy.
• PII: legacy Enron-derived correspondence rows may contain real names and addresses (inherited from the parent; disclosed there too). Claims rows use deterministic pseudonyms.

Citation

@misc{docclass_pilot_2026,
  title  = {Docclass Pilot Sample (KANBAN-084)},
  author = {Lucius-Morningstar},
  year   = {2026},
  url    = {https://huggingface.co/datasets/Lucius-Morningstar/docclass-pilot}
}