Synthetic Form 941-X - Adjusted Employer's Quarterly Federal Tax Return or Claim for Refund Data
Synthetic training data — no real PII, fully coherent identities
Generate synthetic Form 941-X adjusted quarterly payroll returns — 375 fields across five pages of corrections, refund claims, and payroll tax credit adjustments. Every corrected line prints as a corrected amount, an originally reported amount, and a signed difference.
375
Fields per document
5
Pages
Tax
Category
What this document is
Form 941-X, Adjusted Employer's Quarterly Federal Tax Return or Claim for Refund, is how an employer corrects a Form 941 it already filed. The employer names the quarter being corrected and the date the error was discovered, declares whether it is adjusting an underpayment or claiming a refund, then walks a five-page ladder of wage and tax lines. Each line shows the corrected amount, the amount originally reported, the difference between them, and where applicable the resulting tax correction — plus a required written explanation of what went wrong.
Why generate synthetically
Amendment forms are their own document class and they are the class most extraction pipelines get wrong, because the same conceptual line appears three or four times across a row with different meanings and one of those values is signed. They are also the class with the highest fraud and audit stakes: the payroll tax credit lines on this form drove one of the largest wave of improper refund claims in recent IRS history, and any model reviewing those claims needs training data that includes the credit lines populated. No employer will share a real 941-X, because it is by definition an admission of an error attached to a named company.
What makes synthetic data useful
The original and corrected columns are generated as two coherent views of the same quarter with the difference derived from them, so the row arithmetic holds and you can evaluate arithmetic-consistency validation rather than only field extraction. The adjusted-return versus refund-claim election, the certification checkboxes it gates, and the underreported/overreported combination checkbox are drawn as valid combinations rather than independently, so the corpus never contains a legally impossible document. Employer identity is consistent with that employer's Form 941 filings from the same simulated world.
Training challenges
The three- and four-column row structure is the core difficulty: corrected, originally reported, difference, and tax correction are all currency, all similar magnitude, and all within a couple of centimeters of each other, so column assignment errors produce plausible-looking wrong answers that no downstream sanity check catches. The difference column is signed and frequently negative, and small negatives beside large positives are the standard failure mode for currency extraction. Twenty checkboxes encode certification statements whose validity depends on the adjustment type selected three pages earlier — a long-range conditional dependency that local models cannot see. Roughly a dozen certification boxes are legitimately unchecked in almost every document, so a model that predicts unchecked everywhere scores well until it meets the documents that matter. And the free-text explanation area is unconstrained prose in a form otherwise made of numbers, which breaks layout models that assume a fixed field grammar.
Generate synthetic Form 941-X - Adjusted Employer's Quarterly Federal Tax Return or Claim for Refund data
Start with 500 free credits. No credit card required.
Generate NowWho uses this data
Payroll tax credit claim processors and the fraud, audit, and underwriting teams reviewing those claims; tax resolution and amended-return practices automating intake; and payroll platforms building correction workflows on top of extraction. More broadly, any IDP vendor whose amendment-form accuracy has never been measured separately from its return accuracy — the two numbers are rarely close.
Document complexity profile
375 fields across 5 pages, all of them field-bearing: 278 currency amounts arranged as corrected / originally-reported / difference groups, 74 text fields, 20 checkbox targets, and 3 numeric counters. 673 annotation relations, nearly double the field count, because one printed line label governs every column of its row. The binding graph is the third densest we publish — 26 conditional bindings and 314 function calls at a maximum expression depth of 2, with an average binding depth of 0.83 — since almost every printed value is derived from the original and corrected pair rather than sampled.
Key stats from our synthetic corpus
Quantitative characteristics of the Form 941-X - Adjusted Employer's Quarterly Federal Tax Return or Claim for Refund documents our generator produces.
| Metric | Value | Detail |
|---|---|---|
| Adjusted return versus refund claim | 71% / 29% | 71.3% of synthetic 941-X filings are adjusted employment tax returns and 28.7% are refund claims. The election sits on page one and gates which certification checkboxes further down are valid, making it the single most consequential classification on the form. |
| Both under- and overreported amounts | 49% | 48.8% of filings correct underreported and overreported amounts in the same submission, which means both positive and negative differences appear on the same page. That is the configuration most likely to expose sign-handling bugs in an extractor. |
| Median COBRA individuals corrected | 3 | The corrected COBRA individual count has a median of 3 (p25 1, p75 5, max 27) against an original median of 4, with differences running from −3 to +3. These small signed integers exercise a different extraction path than the currency columns. |
| Recovery startup business election | 5% | 5.0% of filings check the recovery-startup-business eligibility box for the employee retention credit. Low-frequency but high-stakes checkboxes like this one are exactly what a naturally collected corpus underrepresents and what a synthetic corpus can guarantee coverage of. |
| Field fill rate | 97% | 364 of the 375 fields carry a value in a typical document. Unlike most large IRS forms, a 941-X is nearly fully populated, because the corrected and originally-reported columns print even for lines that did not change. |
How this document co-occurs with others
Rates at which identities in our corpus that produce a Form 941-X - Adjusted Employer's Quarterly Federal Tax Return or Claim for Refund also produce other documents.
| Correlation | Rate | Detail |
|---|---|---|
| The return being corrected | 100% | Every synthetic 941-X filer also files Form 941. The originally-reported column corresponds to the quarterly return being amended, making the pair a true cross-document correction task rather than an isolated page. |
| Territorial return variant | 100% | Form 941-X carries an explicit checkbox for whether the corrected return was a 941 or a 941-SS. Including the territorial form in the same dataset lets a model learn what that checkbox is actually pointing at. |
| Discrepancy reconciliation | 100% | Employers reconciling an IRS-versus-SSA wage mismatch may file Schedule D as well as a 941-X. The two forms solve the same conceptual problem with different column geometries, which makes them a strong generalization pair. |
All stats above are corpus-derived: they were computed on a local synthetic corpus of 1,000 generated identities produced by SymageDocs' World Simulation Engine. No real employer payroll, credit claim, or refund data was used. `make corpus-stats` regenerates the shipped corpus_stats.json for the site's analytics form set, which does not include this edition; reproduce these figures with `python -m symagedocs.corpus_analysis --corpus-size 1000 --form-id irs_f941x_standard_2024`.
Frequently asked questions
- What data format do synthetic Form 941-X 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 375 fields across all five pages, with the corrected, originally-reported, and difference amounts annotated as separate fields per line rather than one merged row. COCO, YOLO, FUNSD, and BIO/NER exports come from the same job.
- What does a corrected line actually contain?
- Field ids are prefixed correct_, org_, and diff_ for the corrected amount, the originally reported amount, and the difference. Lines that produce a tax consequence carry a fourth cell for the tax correction. Because one printed line label governs the whole row, the form carries far more annotation relations than fields — the relation graph is what encodes the column structure a model needs to learn.
- Are the payroll tax credit and COBRA lines populated?
- Yes. The employee retention and qualified sick and family leave credit lines, the recovery-startup-business election, and the COBRA premium assistance lines all carry values in the corpus. The COBRA individual counts run from 1 to 27 with a median of 3 corrected against 4 originally reported, so the count-correction rows exercise small-integer difference extraction alongside the currency rows.
- Can I use this data commercially?
- Yes. Every employer, wage figure, correction, and credit claim is synthetic, contains no real payroll or refund data, and is licensed for commercial use including model training, benchmarking, and redistribution inside your own products.
- Does it pair with the return it corrects?
- Yes, and it should. Generate Form 941 and Form 941-X for the same identity set and the originally-reported column corresponds to the return being amended, which turns a single-document extraction task into the cross-document reconciliation an actual reviewer performs.