Source Licensed Real-World Document Corpus

1. What we are looking for
Authentic, real-world documents for evaluating data classification models. Documents must be real, with no synthetic or AI-generated content.

Format / attribute Minimum for the full corpus
Structured (CSV, XLSX, JSON, XML) 20%
Unstructured (PDF, DOCX, PPTX, email) 20%
Other formats (e.g., source code) 10%
File size At least 20% in each band: under 100 KB, 100 KB to 1 MB, 1 to 10 MB
Language At least 50% English; multilingual documents welcome
2. Rights and eligibility
• Business documents must come from an organization with the authority to license them.
• Public documents need a license that permits commercial AI/ML use.
• Every document needs provenance and evidence of rights.
3. De-identification requirements
• Personal information must be removed or replaced for every individual named in the document, not only the contributor.
• Replacement values must be realistic surrogates, applied consistently for the same person across documents from one source. Surrogate IDs must not match a real person. No live credentials.
• No synthetic data may be injected into documents.
• Client must never receive original documents or any mapping key. De-identification must be completed before delivery.
• US health documents must meet the HIPAA de-identification standard
4. Geography
• Not eligible: Russia and Ukraine.
• China, Turkey, South Africa and Hong Kong: quality checks must be done in-country, with no personal data leaving before clearance.
• P0, highest priority (47 eligible): Argentina, Australia, Austria, Brazil, Bulgaria, Canada, Chile, China, Colombia, Croatia, Czech Republic, Denmark, Ecuador, Finland, France, Germany, Greece, Hong Kong, Hungary, India, Indonesia, Israel, Japan, Latvia, Lithuania, Luxembourg, Malaysia, Mexico, Netherlands, New Zealand, Norway, Philippines, Poland, Qatar, Saudi Arabia, Singapore, South Africa, South Korea, Spain, Sweden, Switzerland, Taiwan, Thailand, Turkey, UAE, United Kingdom, United States.
• P1, second priority (13): Belgium, Cyprus, Estonia, Greenland, Iceland, Ireland, Italy, Liechtenstein, Malta, Portugal, Romania, Slovakia, Slovenia.
5. Document types
• P0: agreements, bank statement, budget, business plan, customer complaints, customer files, employee disciplinary action, employee insurance, finance, health/medical forms, healthcare, human resources, intellectual property, invoice, legal affairs, loan agreements and offer letters, merger and acquisition files, personal financial information, project documents, quotation, software product development files, source code, statement of work, strategic planning documents, tax documents, wire transfer.
• P1: actuary reports, asset management, business context, completion certificates, compliance policies, construction specifications, control system and SCADA files, credit report, employee pension records, employee stocks and bond records, employment agreement, enterprise risk management, environmental permits, facility permits, factory incident reports, finance policies, financial audit, financial statement, freight documents, garnishment, gifts and entertainment, IT infrastructure and network security documents, information technology, lease deeds, legal agreements, letters of credit, license agreement, manufacturing batch records, marketing collaterals, meeting notes, MoU files, network design files, non-disclosure agreement, OSHA records, paystub, procurement, quality assurance files, resume, safety records, sales and revenue, standard operating procedures and manuals, statement of accounts, work schedules.
• Content-type classifiers: adult/racy/gory images, corporate sabotage, discrimination, harassment, money laundering, profanity, protected material, regulatory collusion, stock manipulation, threat, unauthorized disclosure.
Note: Unstructured documents chats (teams, slack, google meet) the documents must retail full context like chat history with sensitive data elements or emails with all information history and communication.