Datasets:

ZefanCai
/

Open-Jev

114,800 typed rows, 11,600 email requests and 400 whole synthetic families. Original English, Chinese and Turkish emails and labels are CC0-1.0. Source code is MIT. No real mailbox, source emails, IMAP contents, attachments or the community author's private examples were imported.

The pinned community implementation asks two Choice and nine Noul questions, eleven total. Our implementation preserves state shape email.subject, email.from, email.date, email.body and the eleven-head contract with independently authored question wording.

kind distinguishes invoice, payment confirmation, account statement, promotion, newsletter and other. Four evidence Nouls ask about currently owed money, the sender issuing its own charges, billing identifiers and promotion. category distinguishes education, electricity, telecom, banking, airline and other; five independent Nouls also ask about these subjects. Category is unsupervised for nonbill/nonreceipt messages: 4,800 requests retain the runtime question but omit its training label. An application fallback is not ground truth.

Already-paid amounts and advertised prices are not currently owed money. Invoice links without repeated amounts remain invoices; relaying a bill does not make the relay its issuer. Controls distinguish educational services from a student's phone bill, handset retail from telecom service, equipment from electricity supply, bank-mediated payment from a bank's own charges, and known out-of-taxonomy services. Unknown or contradictory grammar is not assigned gold.

Each family contains three languages and nine actions, plus two English invoice taxonomy variants. Reordering categories tests order sensitivity; adding insurance changes the insurance label only when that option exists. No change in confidence is assumed or supervised. Shared unchanged question inputs across taxonomy variants reference their first row. There are 114,800 unique ordered inputs, or 114,400 when candidate order is also disregarded.

Split	Typed rows
train	73,472
calibration	4,018
validation	6,027
test	8,323
ood	22,960

All translations, actions, services and taxonomies for a family share a split. Index modulo ten values 8–9 reserve OOD; other families use stable hashing. OOD reserves fictional identities and moves body paragraphs when possible; two-paragraph messages retain their order. It is not a new grammar for every request. The frozen 10-family probe stays byte-identical in the expanded corpus, including all 290 requests and 2,870 rows in their original split files.

Use only state, question, kind and options as model inputs. Targets, metadata, case reference labels and masks are privileged audit fields. Each Parquet record_json and decompressed raw split retain exact original JSONL bytes and row order, with the same column schema as the existing configs.

No training on this corpus is claimed. A separate real Jev 1.13.0 probe completed 87 held-out requests containing 957 runtime questions and 921 labelled decisions; 36 category questions had no gold. Its 908/921 correct decisions are a small correlated control probe, not full-corpus evaluation or production multilingual accuracy. No provider responses are included in this data upload. The generation manifest records operations at generation time; later probe evaluation is documented separately. Frozen model-training mixtures are unchanged.

From the pinned source checkout, choose a fresh output directory:

python3 -m jev.mailroom_data --output-dir data/mailroom-control-v1-rebuilt --groups 400 --seed 943
python3 -m jev.mailroom_audit --data data/mailroom-control-v1-rebuilt --output mailroom-audit.json

This is a finite original-data task, not a collection of natural private mail, live mailbox automation, or an independently annotated production benchmark.