Synthetic Form 1120-FSC - U.S. Income Tax Return of a Foreign Sales Corporation Data
Synthetic training data — no real PII, fully coherent identities
Generate synthetic Form 1120-FSC foreign sales corporation returns — six pages where every currency amount is split into separate dollar and cent columns. One of the densest forms in the catalog and the reference case for split-column numeric extraction.
606
Fields per document
6
Pages
Tax
Category
What this document is
Form 1120-FSC is the income tax return of a foreign sales corporation, the entity structure used to route export income through a foreign subsidiary. Across six pages it captures foreign trade income, transfer-pricing method elections, shareholder and common-parent information, and a full tax computation. Its distinguishing physical feature is that every monetary amount is printed in two adjacent columns: dollars and cents.
Why generate synthetically
The split dollars/cents column layout appears on a handful of legacy IRS forms and almost nowhere in public training corpora, yet it silently corrupts extraction: a model that reads the two columns as one field produces a value inflated by a hundredfold, and one that reads only the dollar column loses the cents. Synthetic 1120-FSCs give you thousands of labeled instances of that pattern, with the dollar and cent halves annotated as separate ground-truth fields.
What makes synthetic data useful
Every field on this form carries a generated value — no blanks to work around — so a corpus of 1120-FSCs is uniformly dense supervision. Amounts are split into dollar and cent components at generation time rather than being formatted after the fact, which means the ground truth records exactly which digits belong to which column. Entity, shareholder, and common-parent blocks are populated from the same coherent identity, so the multi-party structure of the return holds together.
Training challenges
The dollars/cents split is the headline problem: two numeric columns of very different magnitude sit adjacent under a shared label, and cent values are uniformly distributed across the two-digit range, so there is no statistical shortcut for telling them apart from dollar values. Beyond that, the form is one of the largest we ship, with 942 annotation relations across six pages — more relations than fields, because shared labels govern paired columns. Text fields make up a much larger share here than on other corporate returns, since shareholder, parent, and transfer-pricing election blocks are narrative rather than numeric, so a model tuned for currency-heavy tax pages will under-segment them.
Generate synthetic Form 1120-FSC - U.S. Income Tax Return of a Foreign Sales Corporation data
Start with 500 free credits. No credit card required.
Generate NowWho uses this data
International-tax and transfer-pricing software vendors, trade-finance and export-credit document pipelines, archival and records-digitization providers processing legacy corporate filings, and ML teams building table- and column-extraction benchmarks who need a labeled split-column corpus. Any extractor sold for historical document processing has to survive superseded form layouts, and this is the cleanest test of that.
Document complexity profile
606 fields across 6 pages — 308 currency amounts, 229 text, and 69 checkbox targets — with 942 annotation relations, well above the field count because paired dollar and cent columns share a single printed label. Binding depth is deliberately flat (1 arithmetic binding, 1 function call, maximum depth 1): the challenge on this form is geometric and volumetric, not computational. Every field carries a value, so there are no blank regions to navigate.
Key stats from our synthetic corpus
Quantitative characteristics of the Form 1120-FSC - U.S. Income Tax Return of a Foreign Sales Corporation documents our generator produces.
| Metric | Value | Detail |
|---|---|---|
| Population coverage | 64% | 64% of identities in our 1,000-identity corpus are attached to an employer entity and therefore 1120-FSC eligible. |
| Field fill rate | 100% | Every field on the form carries a generated value in every document. No other form in our catalog offers uniformly dense supervision at this scale. |
| Annotation relations | 942 | 942 label-to-value relations per document against a smaller field count, because dollar and cent columns share a printed label. Relation extraction is the harder half of this form. |
| Text field share | 38% | 229 of the form's fields are text rather than numeric — shareholder, common-parent, and transfer-pricing election blocks. That is a far higher text share than other corporate returns, and it breaks models tuned exclusively for currency grids. |
| Median reported total assets | $23.4M | Median total assets is $23,361,421, with a p25–p75 range of $11.5M to $36.1M. Eight-figure amounts split across dollar and cent columns are exactly where mis-joined column reads produce catastrophic magnitude errors. |
How this document co-occurs with others
Rates at which identities in our corpus that produce a Form 1120-FSC - U.S. Income Tax Return of a Foreign Sales Corporation also produce other documents.
| Correlation | Rate | Detail |
|---|---|---|
| Domestic corporate counterpart | 100% | Every 1120-FSC entity can also produce a standard Form 1120. The pair contrasts a split-column legacy layout against a single-column modern one over comparable corporate content. |
| Other split-column filer (1120-SF) | 100% | Form 1120-SF uses the same dollars/cents split at smaller scale. Generating both gives a split-column training set that spans two very different form sizes. |
| Can file a 7004 extension | 100% | The same entity can produce a Form 7004 extension, whose form code identifies which corporate return is being extended. |
| Files quarterly payroll returns | 100% | Every entity in the corpus is an employer filing Form 941 quarterly, so corporate and payroll filings can be generated as one coherent entity document set. |
| Entity W-9 on file | 100% | Every entity can co-generate a W-9 carrying the name and TIN printed on the return header — the identity anchor for entity-resolution training. |
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, with 1120-FSC figures measured over the entity-eligible subset. No real corporate, shareholder, or trade 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_f1120fsc_standard_2024`.
Frequently asked questions
- What data format do synthetic Form 1120-FSC documents include?
- Each generated identity produces a rendered PDF plus a structured JSON annotation file with bounding boxes and ground-truth values for all 606 fields across six pages, with dollar and cent columns annotated as separate fields. COCO, YOLO, FUNSD, and BIO/NER exports come from the same job.
- How are the dollar and cent columns labeled?
- As two distinct annotated fields sharing one printed label, each with its own bounding box and its own relation back to that label. That is what lets a model learn column identity explicitly instead of inferring it from horizontal offset, and it is why the relation count on this form exceeds its field count.
- Are any fields left blank?
- No. Every field on this form carries a generated value, which makes it the densest uniform supervision in our catalog. If you specifically want blank-field robustness, pair it with the Form 709 or Form 1065 corpora, where large regions are legitimately empty.
- Is this form still filed in the real world?
- The FSC regime was repealed, so the form is legacy. That is precisely what makes it valuable as training data: archival document processing has to handle superseded forms, and a model that only ever sees current-year templates fails on exactly this class of document.
- Can I use this data commercially?
- Yes. All synthetic data is generated from statistical models, contains no real corporate or shareholder data, and is licensed for commercial use including ML model training and benchmarking.