Expert-reviewed finance AI evaluation data
Finance and audit work done by AI, checked the way a senior reviewer would: comparable-company tables, three-statement models, research notes, audit test workpapers and municipal bond summaries, reviewed by CPAs and former investment bankers whose credentials are checked. Every number is traced back to its source, a filing, a statement or an official disclosure, and every error is logged with its correction.
Specification
- Work reviewed
- Comparable-company and precedent-transaction tables, three-statement and DCF models, equity research notes, audit test workpapers and municipal bond summaries
- Sources
- Annual and quarterly reports (10-K, 10-Q and UK annual reports), earnings releases, audited financial statements and bond official statements
- Checks
- Every figure traced to its source page; formulas, periods, units, currency, adjustments and conclusions checked
- Labels
- Per figure: correct, wrong or unsupported, with the right value and its source; per output: score 1–5 and error severity
- Standards
- US GAAP and IFRS for the accounts; audit tests read against the auditing standards they claim to follow
- Reviewers
- CPAs and former investment bankers whose credentials are checked; a second reviewer checks every output
Example of the data
What you see, and the file you get.
| EV / EBITDA | wrong · should be 12.4x | leases left out of net debt |
|---|---|---|
| Revenue FY2025 | correct ✓ | 10-K, p. 48 |
| Audit test | unsupported | sample size not justified |
| output_id | task | item | model_value | correct_value | source | verdict |
|---|---|---|---|---|---|---|
| comps_0014 | trading_comps | EV / EBITDA, LTM | 11.8x | 12.4x | FY2025 10-K, note 9 (leases) | wrong |
| comps_0014 | trading_comps | Revenue, FY2025 | $4,212m | $4,212m | FY2025 10-K, p. 48 | correct |
| model_0031 | three_statement_model | Cash from operations, Q2 2026 | $318m | $296m | Q2 2026 10-Q, cash flow statement | wrong |
| note_0009 | research_note | Gross margin change, year on year | +240 bp | +180 bp | Q2 2026 earnings release, p. 2 | wrong |
| audit_0022 | audit_test | Receivables confirmation sample size | 25 items | Workpaper AR-3 | unsupported | |
| muni_0005 | muni_bond_summary | Debt service coverage, FY2025 | 1.45x | 1.45x | Official statement, p. 112 | correct |
Sample rows. Also as JSON or Excel.
Contact us to get the dataset
Tell us what you need. We walk you through the data, its formats and how it is checked on a short call.
What it is used for
Finance AI benchmarks
Measure how often a model’s numbers survive a professional check.
Fine-tuning and RLHF
Train on reviewers’ corrections, figure by figure, each with its source.
Analyst and audit copilots
Teach assistants to show where every figure comes from.
Error analysis
See which mistakes a model makes most: periods, units, adjustments or plain arithmetic.
Questions
Who checks the outputs?
CPAs and former investment bankers whose credentials are checked, each reviewing the kind of work they have done themselves. A second reviewer checks every output.
What does “every number traced” mean?
Each figure in an output is matched to the page of the filing, statement or official disclosure it came from, or marked unsupported when no source carries it.
Can I see samples first?
Yes. Request samples with the form on this page and we will send them with the full specification.
Tell us what your model needs.
Request samples and the full specification. We reply within a couple of business days.
Request samples