See the data in action
One borrower’s file from our financial dataset: three documents generated from the same ledger, an audit task, and the answer the ledger proves. This customer is a deliberately wrong copy — one error is planted in each document.
correct FAC_34065E1DE0
The statement and the receipt both print FAC_34065E1DE0 — one character differs.
correct 20 Jul 2025 · Purchase · Chaurahvan Fuel Centre · ₹1,964.87 Dr
The purchase lines fall short of “Purchases” in the account summary, and the balance no longer reconciles — both by exactly ₹1,964.87. The purchase’s 2% fee (₹39.30) and its GST (₹7.07) are still printed on 20 Jul.
correct ₹93,545.95
The receipt’s own table of how the payment was applied adds up to ₹93,545.95, with nothing held as credit.
Each finding is a document, page, field and value, checked exactly against the ledger the documents were printed from. A reference detector confirmed all three are findable from the borrower’s own documents.
The same ledger produced the documents and the answer key. Read the answer keys:
Key fact statement ↗Statement ↗Receipt ↗Financial data built for complex AI reasoning
Realistic, interconnected financial documents and account histories, generated from one consistent financial state per account — customers on a credit line, with purchases, repayments, fees and EMIs month by month. Designed for tasks that need multi-document reasoning, numerical accuracy, discrepancy detection and verifiable outcomes.
Key facts, agreements, statements, receipts and letters, with a gold value for every printed field.
Questions that need several documents of the same account to agree before they can be answered.
Balances, fees, tax, EMIs and days past due that must reconcile to the last rupee.
Correct and deliberately wrong copies in pairs, each error with an answer key.
Tasks on a complete customer file — review, reconcile, decide — scored against the ledger.
Question sets with gold answers, the evidence behind each one, and graders that score automatically.
Nine document types per account, from the key fact statement and agreement through statements and receipts to the closure letter and NOC.
Purchases, refunds, repayments — on time, partial, late or bounced — fees with GST, EMI conversion, foreclosure and closure.
Each figure on each page records the ledger entry it came from, and where it sits on the page.
The same account as JSON, as a transaction event feed and as a database export.
Varied layouts, fictional lender brands and number and date formats. The presentation changes; the facts do not.
Credit limit, interest, fees, tax, billing day and due day come from product configuration.
Ground truth you can actually verify
Documents and answers come from the same place — a ledger — so nothing is labelled after the fact.
A fictional borrower and months of account activity.
Every rupee booked in exact decimals and checked.
Documents copy figures from the ledger.
Questions, audits and agent tasks over the documents.
Checked against the ledger, not an opinion.
Every document of an account comes from the same accounting records, so they agree with each other by construction.
The same seed rebuilds the same account, so any scenario and its outcome can be reproduced and audited.
Balances, repayments, charges and account states stay connected from the first statement to the closure letter.
Errors are planted on purpose, one field at a time, and each is confirmed findable from the documents alone.
Correct outcomes are checked against the underlying financial state rather than left to subjective review.
Money held as exact decimals, never floating point
Documents select figures from the ledger and never compute their own
Every account reconciled, validated and replayed from its events before export
Fictional people, lenders and accounts — every page marked as training data
Inspect the data yourself
Real files from the financial dataset. The starter sample needs no sign-up; the larger evaluation pack downloads as soon as you tell us who you are.
Two customers, all nine documents each
One correct, one with an error planted in every document
An answer key for every planted error
File index and notes on how the data was made
Each with the full set of nine documents, from sanction to closure.
Every document of the last three carries one planted, findable error.
The error, the page, the field, the true value, and how to catch it.
Custom data for frontier AI research
We work with AI teams to translate complex data requirements into structured production workflows, combining domain expertise, data generation, validation and quality review.
Lending data built and running today: complete loan documents, the ledgers behind them, planted errors with answer keys, and graded questions. The sample on this page is from this dataset.
Explore the sample →Prompts and expert-quality responses, written to your guidelines.
Held-out, expert-graded sets for measuring a model in one domain.
Ranked responses, rubric scores and worked reasoning from practitioners.
Multi-step tasks with tools, documents and state.
Tasks whose outcomes can be checked programmatically, with graders.
Documents, forms and images from the domain, with structured labels.
Financial services today. Similar data for healthcare, legal and other expert domains, produced to order with a network of domain experts and creators who write the seed material and review what comes out.
Data operations and intelligence for a fintech company in India, and data sourcing, annotation and labelling for a major fraud analytics firm.
Domain, task types, schema, volume and quality bar, agreed with your team.
Domain experts write the seed material — cases, documents and worked answers.
Variations to reach volume, steered for coverage, difficulty and edge cases.
Automated checks first, then expert review; anything that fails is cut.
A sample with review notes, so you can judge quality before volume.
In your schema, with documentation and provenance for each item.
Built for research and production requirements
Delivered in your schema. Financial accounts also come as JSON, a transaction event feed and a database export, alongside the documents.
Accounts are generated in parallel batches, and a 10,000-account generation run has been verified. The same seed rebuilds the same account.
Seed-varied paths — on-time, partial, late and bounced repayments, refunds, reversals, fee waivers, EMI conversion, foreclosure — across varied layouts and formats.
Every account passes accounting checks and an exact replay before export; planted errors are confirmed findable by a reference detector. Custom data adds expert review.
Every printed figure carries a lineage record to the ledger row and field it came from.
Fully synthetic: fictional people, lenders, merchants and accounts, and no real customer data. Licensing terms are agreed per engagement.
Start with the sample on this page or the evaluation pack, then a scoped pilot before any volume commitment.
Building AI that needs better data?
Tell us what you’re training, evaluating, or trying to make verifiable. We’ll explore how to build the right dataset around your requirements.