Ground-truth tax returns for training and evaluating AI
Complete federal and state returns for fictional taxpayers, prepared in professional tax software by working tax professionals.
- 1aWages from Form W-2, box 1215,216Includes $52,400 of restricted stock. The broker’s cost basis already counts it, so it is taxed once.
- 7aCapital gain4,417After a $645 wash-sale loss is disallowed and moved into the basis of the new shares.
- 11aAdjusted gross income219,633
- 15Taxable income203,883
- 16Tax41,712
- 17Alternative minimum tax, from Schedule 27,661From exercising stock options and keeping the shares. Nothing was sold, and regular tax ignores it.
- 23Other taxes, from Schedule 2516Additional Medicare Tax of $348 and net investment income tax of $168.
- 24Total tax49,889
- 33Total payments48,249
- 37Amount owed1,640
Figures are from the prepared return. The taxpayer is fictional.
The free sample set has 20 fictional taxpayers in 20 states, 425 pages of prepared returns and 61 different federal forms and schedules.
One taxpayer per state, each built around a different problem
A farm in Georgia. A military move to Maryland. An Irish company owned from Indiana. A server in Minnesota claiming the new deductions for tips and overtime. Each case was written to make a specific set of forms interact, then prepared the way a filed return is prepared.
A model can fill in a return. Knowing it is right takes a preparer.
A tax return looks like arithmetic and is full of judgment: which form applies, what a document proves, what the taxpayer’s papers leave out.
To train a model on that work, or to score it, you need the return a working preparer would file, and the reasons behind it. That is what we make.
What we make
- Prepared returns
- Complete federal and state returns for fictional taxpayers, prepared in professional tax software and printed as the software prints them.
- Answer keys for evaluation. Targets for training. There are 20 in the free sample set.
- Case files
- The return together with its inputs and its reasoning: a narrative, source documents, every line as data, and a workpaper of the decisions.
- Lets a grader score line by line, and tell a judgment call from an error. The Illinois sample is one.
- Custom cases
- Fictional taxpayers and the papers they would hand over, written so that a chosen set of forms has to interact. They range from a few tax forms to a whole client engagement, with the emails, receipts and bookkeeping reports a firm works from.
- Fills the gaps in your coverage. All 20 sample cases were written this way.
Five places a model goes wrong on one return
The Illinois sample is a software engineer with restricted stock, stock options and one badly timed sale. Nothing about it is exotic, and every step needs a decision.
Restricted stock income is taxed once
Box 1 of the W-2 already includes $52,400 from two vests, and the broker’s reported basis includes it too. The sales go on Form 8949 as reported. A model that adds the income again, or “fixes” a basis that is already right, gets the gain wrong.
Form W-2 box 14, Form 8949
A wash sale moves the loss; it does not erase it
One hundred fifty shares were sold on August 20 at a $645 loss, and new shares vested on September 15, inside the 30-day window. The loss is disallowed and added to the basis of the new shares. Form 8949 carries code W and a $645 adjustment, so that sale nets to zero and short-term gain comes to $928.
Form 8949 Part I, Schedule D line 7
Exercising options can cost tax with no sale
The spread on 1,600 options is $59,840: $59.40 less $22.00, times 1,600 shares. Regular tax ignores it because the shares were kept. The alternative minimum tax does not, and Form 6251 produces $7,661.
Form 6251 line 2i, Schedule 2 line 2
Two surtaxes use two different measures
Additional Medicare Tax runs on Medicare wages of $238,716, which include the 401(k) deferral, and comes to $348. Net investment income tax runs on the smaller of investment income and income over $200,000, and comes to $168. The employer withheld $349 toward the first, and that goes on Form 1040, line 25c.
Form 8959, Form 8960
The state return follows its own rules
Illinois starts from federal adjusted gross income, allows one exemption and applies a flat 4.95%. Chicago has no city income tax. The result is $10,731 of tax and $78 owed after withholding.
IL-1040, Schedule IL-WIT
How a case is made
The work is done by a team of 16 preparers, 4 reviewers and 2 people on quality assurance.
Write
We invent a taxpayer and the documents they would hand over, built around the forms you want to test. Or you send your own.
Prepare
A preparer builds the federal and state returns in professional tax software, the way a filed return is built.
Review
Every return goes through two layers of review against the source documents. Quality assurance then checks that the files are complete and intact.
Deliver
You receive the files under your own IDs. If something is wrong, we fix it and reissue it.
Invented taxpayers, real experience
We adhere entirely to Section 7216 and never reuse any client data.
- What is invented. Every name, address, identification number, employer, account and figure in a case is made up for that case.
- What is real. The experience. We may use kinds of situations we have seen before, such as a prenuptial agreement in Washington that allocates assets, income and liabilities between spouses, or a taxpayer who lives in another country and works remotely through an LLC in the United States.
- Coincidence. Cases are written to look real, so a made-up name or number can match a real one by chance.
- Your files stay yours. Files you send us are used for your project and nothing else.
Start with the sample
See how one return comes together, then ask for all 20. Tell us what you are training or evaluating and we will reply with scope, timing and price.