Synthetic W-2 Wage and Tax Statement Data
Synthetic training data — no real PII, fully coherent identities
Generate synthetic 2023 vertical-stack W-2 sheets — four full-width copies stacked top to bottom on one page. The oldest of the three stacked revisions, with the 2023 Social Security wage base, for testing whether a W-2 model holds up on a two-year-old document.
144
Fields per document
1
Page
Tax
Category
What this document is
The vertical-stack W-2 prints four full-width copies of the same Wage and Tax Statement one above the other down a single page, rather than in a two-by-two grid. This is the 2023 revision. It is the layout that dominates continuous-feed and laser payroll print streams, and it is the layout a scanner most often mangles, because a page-height stack of four near-identical bands is easy to mis-crop and easy to mis-order.
Why generate synthetically
A 2023 statement is not a historical curiosity — it is the second year in every two-year income lookback performed during 2025 and the prior-prior year in any three-year self-employment review. Yet almost no training corpus contains a balanced sample of it, because corpora are built from whatever was current when they were built. Generating a prior revision with full ground truth gives you a genuine held-out year rather than a random split of one year, and it does so without any of the sourcing problems that make real W-2s unshareable in the first place.
What makes synthetic data useful
Each sheet is one simulated employee's 2023 statement replicated down the page with identical values in four distinct coordinate bands. Box 3 Social Security wages cap at the 2023 wage base of $160,200 while Box 5 Medicare wages continue past it, which places this revision one rung below 2024 and two below 2025 on the wage-base ladder. Holding the simulated population fixed across all three stacked revisions means the only thing that changes between them is the document, so any accuracy difference you measure is attributable to the revision rather than to the sample.
Training challenges
Stacking four full-width copies vertically produces a page where every label appears four times at the same horizontal position and four different vertical offsets. That is a harder confusion than the quadrant layout: horizontal position, which is normally the strongest disambiguating signal for a wide form, carries no information here at all. A model must use vertical banding alone to decide which copy a value belongs to, and because all four copies carry identical values the error is invisible to any evaluation that scores values without scoring copy assignment. Each band is a quarter of a page tall and full width, so boxes are wide and short — an aspect ratio that off-the-shelf detectors handle poorly.
Generate synthetic W-2 Wage and Tax Statement data
Start with 500 free credits. No credit card required.
Generate NowWho uses this data
Mortgage and consumer-lending platforms performing two-year income lookbacks, payroll-service scanning operations that ingest continuous-feed print output, and document-AI teams that need a held-out revision rather than a held-out random split. Fraud teams use prior-year stacks to train on the altered-historical-document case, which is where most W-2 fabrication actually occurs.
Document complexity profile
144 fields on a single page: 72 currency amounts, 52 text, 12 checkbox targets, and 4 Social Security number and 4 employer identification number typed fields, arranged as one 36-field statement stacked four times down the page. 108 annotation relations, 12 conditional bindings, 16 function calls, maximum expression depth 2. Every copy sits at the same horizontal offset, which removes horizontal position as a disambiguating signal.
Key stats from our synthetic corpus
Quantitative characteristics of the W-2 Wage and Tax Statement documents our generator produces.
| Metric | Value | Detail |
|---|---|---|
| 2023 Social Security wage base | $160,200 | Box 3 Social Security wages cap at $160,200 while Box 5 Medicare wages run to $316,900. This is the lowest rung of the three-year stacked ladder and the value a current-year-trained validator is most likely to reject as out of range. |
| Copies stacked per page | 4 | Four full-width statements stacked vertically at identical horizontal offsets. Because horizontal position carries no disambiguating information, copy assignment must come from vertical banding alone — a materially harder problem than the two-by-two quadrant sheet. |
| Median Box 1 wages | $45,600 | Across 641 eligible synthetic identities the median Box 1 wage is $45,600, p25–p75 $27,200 to $74,400, spanning $4,000 to $316,900. Identical across the three stacked revisions by construction, so year-over-year differences come only from the document. |
| Blank field rate | 42% | 60 of the annotated fields render blank on a typical sheet — allocated tips, dependent care, nonqualified plans, Box 12c and 12d, the statutory-employee and sick-pay checkboxes, and the local wage block, repeated across four bands. All of them are annotated. |
| Retirement plan participation | 57% | 57.1% of sheets check the Box 13 retirement plan box, with a matching Box 12a elective-deferral code — 71% D, 17% E, 12% G. In a full-width band a quarter of a page tall, the Box 13 checkboxes are wide-and-short targets that generic detectors systematically miss. |
How this document co-occurs with others
Rates at which identities in our corpus that produce a W-2 Wage and Tax Statement also produce other documents.
| Correlation | Rate | Detail |
|---|---|---|
| Next rung of the wage-base ladder | 100% | The 2024 stack for the same employee, identical in every respect except a Social Security cap of $168,600. Two revisions differing in exactly one value is as controlled a comparison as document AI gets. |
| Quadrant layout, same content | 100% | The four-up sheet carries the same wage semantics in a two-by-two grid. Comparing accuracy on the stack against accuracy on the grid isolates how much of a model's field association depends on horizontal layout. |
| Employee's personal return | 100% | The wages on the statement flow to the employee's Form 1040. Prior-year W-2 plus current-year return is the standard bundle in amended-return and back-tax processing. |
| Issuing employer's quarterly returns | 100% | Every synthetic W-2 is issued by an employer that also files Form 941, so an employer's W-2 wages aggregate against its quarterly wage totals — the canonical payroll reconciliation task. |
Wage and rate figures above are corpus-derived: they were computed over the 641 employment-eligible identities inside a local synthetic corpus of 1,000 identities generated by SymageDocs' World Simulation Engine at seed 20260421. Field, type and relation counts come from the shipped 2023 vertical-stack W-2 definition in the SymageDocs form library. No real employee, employer, or payroll data was used at any stage.
Frequently asked questions
- What data format do synthetic vertical-stack W-2 documents include?
- Each generated identity produces a rendered PDF plus a structured JSON annotation file with bounding boxes, field types, and ground-truth values for all 144 fields on the sheet — 36 fields per copy across four stacked copies — plus 108 label-to-value relations. Each annotation records which band it belongs to. COCO, YOLO, FUNSD, and BIO/NER exports come from the same job.
- How does the stack differ from the four-up quadrant sheet?
- Same content, different spatial arrangement. The four-up sheet places copies in a two-by-two grid, so horizontal position distinguishes them; the stack places them one above the other at identical horizontal positions, so only vertical banding does. Running the same model against both is a clean ablation of how much of its field association is really coming from horizontal layout priors.
- What was the 2023 Social Security wage base?
- $160,200. Box 3 caps there while Box 5 Medicare wages continue uncapped. Together with the 2024 sheet at $168,600 and the 2025 sheet at $176,100, the three stacked revisions give you a three-rung wage-base ladder over an otherwise identical document — the cheapest available test of whether a validator treats the cap as a year-indexed parameter or a constant.
- Is the wage distribution the same across the three stacked years?
- Yes, deliberately. The simulated population is held constant so that the only variable between the 2023, 2024 and 2025 stacks is the revision itself. Median Box 1 wages are $45,600 on all three; only the Social Security cap moves.
- Can I use this data commercially?
- Yes. Every Social Security number, name, address, employer and wage figure is synthetic, contains no real personal data, and is licensed for commercial use including model training, benchmarking, and redistribution inside your own products.