The v1.0.0 head read a mortgage statement as Form 1098 at 0.86 because the
data had no account-statement classes, one template style and no hard
negatives. v1.1.0 adds 7,271 programmatically generated rows on top of the
unchanged v1.0.0 rows, label IDs and splits.
- 12 new fine labels (IDs 83-94) and 7 new families (IDs 42-48), including
MORTGAGE_STATEMENT, PAYSTUB, CONSOLIDATED_1099 and OTHER_TAX_FORM
- Realistic multi-layout rows for 49 existing labels: real IRS box labels,
title top/bottom/removed, OCR noise, compact snippets, raw CSV exports
- Hard negatives for statements that mention tax forms; UNKNOWN_DOCUMENT
broadened to 26 kinds
- Held-out layouts per label and held-out kinds for OTHER/UNKNOWN; zero
cross-split group, exact or digit-normalized leaks
- Header-masked v1.1 view, 130-text external probe set (dev and blind),
generator, TF-IDF baseline and frozen-MiniLM probe scripts and reports
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Add Finipher US Tax document classifier dataset v1.0.0
Upload 6,640 AI-labeled synthetic document-classification examples covering 83 detailed document types and 42 broader families.
Includes W-2, 1099 subtypes, 1098, gift and donation records, statements, receipts, instructions, and unknown documents.
Provides string and integer labels, grouped train/validation/test splits, CSV, JSONL and Parquet files, label mappings, provenance, and benchmark reports.
Intended for classifier training and research. Real-document production accuracy has not been validated.
