Skip to content
Outrich Data

Free template

Data card template

A data card is the short description that travels with a dataset. This free template covers what an AI lab asks first: source, record types, schema, volume, time span, processing, de-identification method, known gaps, history and the license scope on offer. Copy it, fill it from measurements and date each version.

Data card: [dataset name]

Version: [number]

Date: [YYYY-MM-DD]

Prepared by: [name and role]

1. Summary

[Two or three sentences. What the dataset is, where it comes from and what kind of work it records.]

2. Source

  • Supplier profile: [industry and size band, no company name unless agreed]
  • Source systems: [for example helpdesk, chat tool, document store]
  • How records were created: [by employees in the course of work, by customers, by automated systems]
  • Export date: [YYYY-MM-DD]
  • Export method: [vendor export, API, database query]

3. Contents

  • Record type and unit: [for example one resolved support conversation]
  • Record count: [number]
  • Estimated tokens in content fields: [number and tokenizer, or state that it is an estimate]
  • Time span: [earliest record date] to [latest record date]
  • Languages: [list with approximate share]
  • Regions covered: [list]
  • File format: [for example JSONL, UTF-8, one object per line]
  • Number of files and total size: [number, size]

4. Schema

FieldTypeDescriptionFill rate
[id][string][hashed record identifier][percent]
[created_at][ISO 8601 date][when the record was opened][percent]
[messages][list of objects][ordered turns with role and content][percent]
[outcome][string][final status of the record][percent]

5. Records by period

PeriodRecord count
[year][number]
[year][number]

6. Processing

  • Filters applied: [what was excluded and why]
  • Duplicate rule and rows removed: [rule, number]
  • Fields dropped: [list]
  • Fields hashed: [list, method, where the salt is held]
  • Format conversion: [from, to, tool]

7. De-identification

  • Standard relied on, if any: [for example HIPAA Safe Harbor, or none claimed]
  • Identifier classes removed: [names, email addresses, phone numbers, account numbers, addresses, other]
  • Method: [tools used, pattern scan, manual review]
  • Testing: [sample size, how drawn, result]
  • Performed by: [team or provider]
  • Known residual risk: [for example free text may contain indirect references]

8. Known gaps and limits

  • [Periods with missing or thin data]
  • [Record types excluded]
  • [Customer accounts removed after rights review]
  • [Quality issues a user should know about]

9. History

  • Ever public, in whole or part: [yes or no, with detail]
  • Previously licensed: [no, or to whom in general terms and on what basis]
  • Contains synthetic or machine-generated records: [yes or no, with detail]

10. Rights and license scope offered

  • Ownership basis: [created by the supplier's employees under assignment]
  • Title report available: [yes or no, date]
  • Uses offered: [pre-training, fine-tuning, evaluation, retrieval]
  • Exclusivity: [non-exclusive, or terms]
  • Term: [fixed term or perpetual]
  • Refresh available: [no, or schedule]

11. Integrity

  • Checksum manifest: [file name, hash algorithm SHA-256]
  • Sample available under NDA: [yes or no, size]

Find out if your data qualifies.

Eleven questions, about five minutes, no call required. You get an answer on screen and a written follow-up.

esc