finipher/finipher-us-tax-document-classifier

Finipher Document Classifier Dataset

Version 1.1.0. A flat supervised text-classification dataset with 13,911 unique synthetic observations, 95 fine labels and 49 family labels. It adds 7,271 programmatically generated rows to the unchanged 6,640 rows of v1.0.0.

Why v1.1.0

The v1.0.0 head (tax-document-head-20261017-v1) read a mortgage statement as Form 1098 at 0.86 confidence. The cause was the data, not the encoder:

• No account-statement classes. Mortgage, loan, brokerage, retirement and HSA statements, paystubs, insurance documents, crypto reports, property tax bills and P&L statements had no label, so the softmax had to pick the nearest tax form.
• One template style. Every v1.0 row is about 35 words in a "Form X: Title / Calendar year / fields" pattern. Real extracted text is longer, laid out in many ways and often noisy.
• No hard negatives. Nothing taught the head that a statement which mentions Form 1098 is not Form 1098.
• No route for unseen forms. A real tax form outside the vocabulary (SSA-1099, 3921, a state return) could only land on a wrong known form.

What changed

Area	v1.0.0	v1.1.0
Rows	6,640	13,911 (v1.0 rows kept as is)
Fine / family labels	83 / 42	95 / 49 (new IDs appended; v1.0 IDs unchanged)
Mean words per row	34.5	84.4 for new rows (p90 151)
Layouts	1 template family	stacked, inline, two-column, boxed, flat (pdftotext-like), markdown (Docling-like), CSV export
Title position	always first	top 70%, bottom 15%, removed 15%
OCR noise	none	about 20% of new rows (glyph swaps, merged words, $→S)
Label source	single AI annotation	programmatic ground truth from the generator (new rows)

New fine labels (IDs 83–94): MORTGAGE_STATEMENT, LOAN_STATEMENT, BROKERAGE_STATEMENT, RETIREMENT_ACCOUNT_STATEMENT, HSA_ACCOUNT_STATEMENT, PAYSTUB, INSURANCE_DOCUMENT, CRYPTO_TAX_REPORT, PROPERTY_TAX_BILL, PROFIT_AND_LOSS_STATEMENT, CONSOLIDATED_1099 and OTHER_TAX_FORM.

New families (IDs 42–48): LOAN_STATEMENT, INVESTMENT_STATEMENT, PAYSTUB, INSURANCE_DOCUMENT, CRYPTO_TAX_REPORT, PROPERTY_TAX_BILL and OTHER_TAX_FORM. PROFIT_AND_LOSS_STATEMENT maps to BUSINESS_RECORD and CONSOLIDATED_1099 maps to 1099.

Realistic rows for existing labels. Real IRS box labels, OMB numbers, copy lines and payer/recipient blocks, for:

• W-2, W-2C and W-2G
• 1099-INT/DIV/B/R/NEC/MISC/G/K/SA/Q/S/C/DA
• 1098, 1098-T/E/C
• 5498, 5498-SA/ESA
• 1095-A/B/C
• 1040, 1040-X, 1040-ES
• the three K-1s, W-9, 8949, 8283, Schedule C and Schedule E

Statement-style rows were added for:

• bank and card statements
• giving statements and donation appeals
• charity event, medical and purchase receipts
• estimated-tax payment confirmations and tax instructions
• unknown documents

A quarter of form rows are compact snippets with a title line and one to four boxes. A quarter of statement rows are raw account exports: a header row and data rows, with no title.

Hard negatives. These pairs share vocabulary by design:

Document	Easily confused with
Mortgage statement (and escrow analysis, payoff letter)	1098
Bank interest summary	1099-INT
Brokerage statement	1099-DIV, 1099-B, consolidated 1099
Retirement statement	5498, 1099-R
HSA statement	5498-SA, 1099-SA
Paystub, including the final one of the year	W-2
Student loan statement	1098-E
Crypto gain/loss report	1099-DA, 8949
Property tax bill	1098 box 10
EOB or declarations page	1095
P&L	Schedule C
Card year-end summary	1099-K
Payment confirmation	1040-ES voucher

About half of the statements also carry a line such as "Your Form 1098 will be available by January 31". The label is what the document is, not which forms it mentions.

UNKNOWN_DOCUMENT now covers 26 everyday kinds instead of one "illegible fragment" template, for example: utility and phone bills, a lease, vehicle registration, a jury summons, an offer letter, a credit report, a bank ad, a tax-prep flyer, HOA notices, parking tickets, recipes, contact and gradebook CSVs, timesheets and sensor logs.

OTHER_TAX_FORM holds 16 real forms outside the vocabulary, for example SSA-1099, RRB-1099, 1099-LTC, 1099-PATR, 1042-S, 3921, 3922, 2439, 8606, the ABLE forms, and IL-1040 / CA 540 / NY IT-201.

Files and splits

Split	Rows	v1.0 rows	v1.1 rows
Train	8,864	3,984	4,880
Validation	2,516	1,328	1,188
Test	2,531	1,328	1,203

• Held-out layouts. For every generated label, one of its six layouts appears only in test and another only in validation. The test measures new-layout generalization, not memorized layout.
• Held-out kinds. OTHER_TAX_FORM and UNKNOWN_DOCUMENT are split by form or kind. Test contains 3922 and 1098-Q, and school newsletters, insurance letters and résumés, none of which appear in training.
• Leak checks. 0 cross-split group leaks, 0 exact-text leaks and 0 digit-normalized leaks (reports/data-quality.json). Within-generator near-duplicates were removed: 39 instruction rows and 10 unknown rows.
• Configs. data/<configuration>/<split>.{parquet,jsonl,csv} holds the four configurations as before. The *_ids Parquet files embed Hugging Face ClassLabel names.
• New metadata columns: layout, title_mode, ocr_noise and sub_kind. Use them for slicing only, never as model input.
• Two-column CSVs. simple_csv/<target>/<split>.csv has exactly text,label.
• robustness_views/test_header_masked_v1_1.{parquet,jsonl}: every generated test row re-rendered with all title lines removed. It is a paired stress view, not an independent split. The v1.0 view test_header_masked.parquet is unchanged.
• robustness_views/external_probes_v1_1.jsonl: 130 texts outside the generator, each with an accept list:
  • 66 Finipher platform test fixtures (synthetic PDFs and CSVs) plus 31 hand-written snippets, marked dev_used_to_find_gaps;
  • 33 hand-written snippets written before the final evaluation was run, marked blind_written_before_eval.
• scripts/v1_1/: the generator (specs.py, render.py, docs.py, build.py), baseline.py and minilm_probe.py.
  • python scripts/v1_1/build.py rebuilds the release deterministically, reading v1.0.0 from commit cad7079.

Measured results

Character TF-IDF + logistic regression (reports/baseline-metrics.json, train-only fit):

Target	Test acc.	Test acc., v1.1 rows	Header-masked v1.1
Fine type	99.6%	99.1%	97.9%
Family	99.7%	99.3%	97.9%

Frozen all-MiniLM-L6-v2 with a linear head, using the production settings (mean pooling, L2 normalization, 128 tokens) (reports/minilm-probe-metrics.json). No checkpoint is shipped.

Evaluation	v1.0 production head	Linear head refit on v1.1 train
v1.0 test rows	100%	100%
v1.1 test rows, existing labels	58.1%	97.2% (all v1.1 test rows)
v1.1 header-masked test	—	94.1%
External probes, dev (97)	40 / 97	80 / 97
External probes, blind (33)	14 / 33	29 / 33
Blind probes with confidence ≥ 0.7	5 / 6 correct	22 / 22 correct

The mortgage statement that started this release goes from 1098 (0.86) to MORTGAGE_STATEMENT (0.99).

Remaining errors are concentrated where text alone is ambiguous:

• Loan and mortgage CSV exports that share an identical header with no type word.
• IRA, HSA and 401(k) exports with the same columns as brokerage holdings.
• A Dutch-language bank export.
• 1099-MISC vs 1099-NEC once the title is removed.

The first two are still correct at the family level. Use the file name or account metadata as an extra signal there rather than expecting the text model to guess.

Recommended use

1. Fine-tune or refit with abstention. Route anything below a validated confidence threshold, or predicted as UNKNOWN_DOCUMENT or OTHER_TAX_FORM, to review instead of a bucket. On the blind probes every error had confidence < 0.6.
2. Calibrate on real documents. Synthetic test accuracy is an upper bound. Before relying on the threshold, measure it on consented, independently labeled real documents.
3. Keep IDs fixed. IDs 0–82 and 0–41 are the v1.0 IDs. A v1.0 checkpoint loads, but it cannot output the new labels until it is retrained.
4. Mind the token budget. New rows average about 84 words. At 128 tokens, rows whose title sits at the bottom are classified from content alone, which is intended and mirrors real extraction order.

Scope, provenance and limits

All rows are synthetic. Names, institutions, amounts and IDs are invented. No private taxpayer documents, production extraction or production model outputs were used. v1.0.0 rows keep their single-AI annotation and lineage. v1.1.0 rows carry generator ground truth and fgen-* lineage in metadata/provenance.jsonl, including generator hash and seed key.

Programmatic labels are exact with respect to the generator, but the generator is a model of real documents, not real documents. The dataset is US-federal-first and English only. High synthetic scores do not establish real-world, OCR or unseen-layout accuracy, and an UNKNOWN_DOCUMENT or OTHER_TAX_FORM class does not guarantee rejection of every unseen document. A recognized form does not establish receipt, taxability or completeness. Not production-validated real-document data. No independent CPA review has been done.

Original synthetic text and annotations use CC BY 4.0. No upload has been performed.