Synthetic W-2 Wage and Tax Statement Data
Synthetic training data — no real PII, fully coherent identities
Generate synthetic 2025 W-2 documents in the full eleven-page official IRS packet, where every field lives on one interior page and the other ten are copies and instructions. The page-routing benchmark for multi-page document pipelines.
45
Fields per document
11
Pages
Tax
Category
What this document is
This is the W-2 as the IRS actually publishes it: an eleven-page packet containing Copy A, the employer and employee copies, the state and local copies, and the accompanying instruction pages. It is what a filer downloads, what a scanner sees when someone uploads the whole download, and what almost no document-AI corpus contains, because corpora are built by cropping the one interesting page out and discarding the rest.
Why generate synthetically
Discarding the rest is exactly what makes production pipelines fail. In the packet the data lives on a single interior page — not the first one — so any heuristic that classifies a document from page one, or that assumes the useful content is at the front, gets the wrong answer on a document that a human recognises instantly. Training on cropped pages teaches a model to extract from a page it has already been handed; the packet forces it to find the page first. Ground truth for that routing decision only exists if the document is generated whole.
What makes synthetic data useful
Each generated packet carries one simulated employee's 2025 wage data on its data page, with Box 3 Social Security wages capped at the 2025 wage base of $176,100 and Box 5 Medicare wages uncapped, and with the remaining ten pages present and empty exactly as published. The annotation file states which page carries fields and, by omission, that the other ten carry none — a supervised negative for every page-classification and page-selection stage in the pipeline. Because the same synthetic employees also generate the half-page, four-up and stacked layouts, the packet can be evaluated against a known-correct extraction of the same wage data from an easier document.
Training challenges
Three failure modes show up here that a single-page corpus cannot expose. Page routing: the data page is an interior page, so front-page classification fails. False positives on empty pages: ten pages of copy headers and instruction text look enough like a form to draw spurious extractions from a model that has never been shown a page whose correct output is nothing. And cost: pipelines that OCR every page of every upload spend eleven pages of compute to read one, which is a throughput problem long before it is an accuracy problem. Roughly half the fields on the data page are legitimately blank on top of that.
Generate synthetic W-2 Wage and Tax Statement data
Start with 500 free credits. No credit card required.
Generate NowWho uses this data
Document-management and IDP vendors whose pipelines must route a multi-page upload before extracting from it, income-verification platforms that receive whatever PDF the borrower downloaded, and teams sizing OCR cost per document rather than per page. Fraud and payment-integrity teams use the packet because a substituted interior page is the cheapest W-2 forgery there is.
Document complexity profile
45 fields across an eleven-page document, all of them on a single interior page: currency, text, checkbox, Social Security number and employer identification number types, with one label-to-value relation per field. 4 conditional bindings and 5 function calls, maximum expression depth 2. Ten of the eleven pages carry no fields at all — the highest ratio of correctly-empty pages to data pages in the catalog.
Key stats from our synthetic corpus
Quantitative characteristics of the W-2 Wage and Tax Statement documents our generator produces.
| Metric | Value | Detail |
|---|---|---|
| Pages carrying fields | 1 of 11 | All fields sit on one interior page; the other ten are the remaining copies and the instruction pages. Ground truth for those ten is an explicit empty set, which turns page routing into something you can score instead of assume. |
| 2025 Social Security wage base | $176,100 | Box 3 Social Security wages cap at $176,100 while Box 5 Medicare wages continue to $316,900. The packet and the stacked and half-page 2025 layouts all share this cap, so a validator can be tested for year-awareness independently of layout. |
| Blank field rate | 51% | Just over half the fields on the data page render blank for a typical employee. Combined with ten empty pages, the packet is the catalog's strongest test of whether a model can say nothing when nothing is there. |
| Median Box 1 wages | $43,027 | Across 641 eligible synthetic identities the median Box 1 wage is $43,027, p25–p75 $25,764 to $72,000, spanning $4,000 to $315,300. This layout computes Box 1 as federal taxable wages — gross pay minus pre-tax retirement deferrals — so its median sits below the $45,600 gross-wage figure quoted for the four-copy and stack layouts; the underlying identity population is the same. |
| Retirement plan participation | 57% | 57.1% of packets check the Box 13 retirement plan box with a matching Box 12a elective-deferral code — 71% D, 17% E, 12% G. Unlike the multi-copy sheets, the packet shows each of these exactly once, so there is no redundant second look to recover from a 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 year's packet | 100% | The 2026 packet raises the Social Security cap and splits Box 14 into 14a and 14b for Treasury Tipped Occupation Codes. Same eleven-page structure, so the pair isolates the regulatory change from the document structure. |
| Same data, one page | 100% | The half-page single copy carries the same 2025 wage data with no routing decision. Running both is how you separate extraction accuracy from document-navigation accuracy. |
| Same year, stacked layout | 100% | The stacked sheet puts four copies of the same year's statement on one page. Packet and stack are opposite ends of the multi-page problem: ten empty pages there, four crowded bands here. |
| Employee's personal return | 100% | The wages on the statement flow to the employee's Form 1040. Uploading a whole downloaded packet alongside a return is the exact multi-document, multi-page bundle consumer tax platforms have to survive. |
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, page, type and relation counts come from the shipped 2025 W-2 packet 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 W-2 packet documents include?
- Each generated identity produces a rendered eleven-page PDF plus a structured JSON annotation file with bounding boxes, field types, and ground-truth values for all 45 fields, together with one label-to-value relation per field. Every annotation is tagged with its page, and the ten pages without fields are part of the same document. COCO, YOLO, FUNSD, and BIO/NER exports come from the same job.
- Which page of the packet carries the data?
- An interior page — the second. The first page is the Copy A sample, and the pages after the data page are the remaining copies and the instruction pages. That ordering is the point: a pipeline that classifies or extracts from page one sees the wrong page, and only a document generated whole can teach it otherwise.
- How is this different from the half-page single-copy W-2?
- Same wage semantics, radically different document. The single-copy layout is one statement on one page with no routing decision to make; the packet wraps the same content in ten pages of context. Evaluating a model on both is how you separate its extraction accuracy from its document-navigation accuracy — two things that a single-page benchmark reports as one number.
- Why are so many fields blank?
- Because they are blank on real statements. Just over half the fields on the data page render empty for a typical employee — allocated tips, dependent care benefits, nonqualified plans, the third and fourth Box 12 slots, the statutory-employee and third-party sick-pay checkboxes, and the local wage block. All of them are annotated, so an extractor that invents plausible values can be measured rather than rewarded.
- 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.