DeepData/Datasets/Long-Document Reasoning
Data line

Specialist tasks built from real PDF and Word documents in finance, law and healthcare. Each requires cross-file retrieval, evidence extraction, calculation and reasoning. We prioritise tasks current models get wrong, but whose answers can be verified.

Scale10K+ tasks
DomainsFinance · Legal · Healthcare
SourceReal PDF & Word
AnswersVerifiable
01Why this data

What makes it worth training on.

Expert-level difficulty

Selected from tasks where current frontier models fall into traps.

Graduate & competition math: 1M+ problems →
Verified, not just generated

Every task carries a verifiable reference answer and its evidence.

98%+ sampled accuracy · 97%+ customer-validated →
Built to your specification

Domains and document types chosen to match your target use.

Scope → pilot → acceptance → scale →
02Sample

Samples on request.

Long-document tasks are built on real professional documents, so samples are shared directly under NDA rather than publicly.

Request a sample →
03Coverage

How it is sliced.

Finance

Extract data from filings, reports, confirmations, contracts and invoices; calculate metrics, assess risk, analyse valuation.

Legal

Retrieve across statutes and regulations; interpret provisions and reach well-grounded conclusions.

Healthcare

Read specialist long-form documents and synthesise evidence to answer professional questions.

Selection

Real documents, verifiable answers and observed model errors: built for targeted RL, usable as a hard benchmark.

04Schema

Record format.

FieldTypeDescription
documentsarraySource files parsed to Markdown, with page references.
questionstringSpecialist question requiring cross-document reasoning.
answerstringVerifiable reference answer.
evidencearraySpans and pages supporting the answer.
reasoningstringExpert reasoning path, including calculations.
domain / difficultyenumFinance · Legal · Healthcare; difficulty grade.
05Use

Built for.

  • Expert long-context RLTargeted training on the failure modes of current models.
  • BenchmarkingMeasure professional reasoning where answers are verifiable.
  • Domain agentsGround finance, legal and healthcare agents in real documents.
Contact

Request a pilot slice.

Tell us the long-document reasoning slice you need. We will send a pilot sample cut to your specification for your own evaluation.