# Datasheet: [dataset name] ## 1. Identification | Field | Value | |---|---| | Dataset id | [e.g. ds-2026-014] | | Version | [1.0] | | Date of this datasheet | [YYYY-MM-DD] | | Supplier | [legal name of the party delivering the data] | | Supplier contact | [role and email address] | | Licence reference | [contract or licence id] | | Schema | [e.g. aidb.conversation/1.0] | ## 2. Purpose and permitted use - Why the dataset was assembled: [the buyer's spec in one or two sentences] - Uses the licence permits: [training / fine-tuning / evaluation / RL environments / other] - Uses the licence excludes: [e.g. re-identification, resale, use outside named models] - Uses the supplier advises against, even where permitted: [e.g. inferring traits of individuals] ## 3. Composition | Field | Value | |---|---| | Data types | [workplace chat / email / support tickets / source code / documents / CRM records] | | One record is | [a thread, a ticket, a repository, a document] | | Records | [count] | | Tokens | [count, and the tokenizer used to count them] | | Languages | [BCP 47 tags with share of tokens, e.g. nl 62%, en 31%, de 7%] | | Period covered | [first and last month, before the date offset] | | Sectors and functions | [e.g. wholesale; sales, purchasing, customer service] | | Number of licensing organisations | [count, or "1"] | - What a typical record looks like: [short description, or point to the sample file] - What was deliberately left out: [e.g. HR mailboxes, direct messages, attachments] - Known gaps and skews: [e.g. one team dominates, a year is missing, mostly one product line] - Does the dataset contain personal data after de-identification? [yes / no, and which residual categories] - Does it contain copyright-protected works of third parties? [yes / no / unknown, and how handled] ## 4. Provenance and rights - Who holds the rights to the source records: [organisation type, not name if confidential] - Basis on which the supplier may license them: [e.g. licence from the rightsholder, dated YYYY-MM-DD] - Export route: [e.g. workspace admin export, eDiscovery export, helpdesk export] - Who ran the export and when: [role, date] - Third-party content found and its treatment: [e.g. open-source code under MIT kept, GPL files removed] - Chain of custody: [where the raw export was stored, who had access, when it was deleted] ## 5. Selection and processing - Selection rules: [which records were included and why] - Filtering: [bots, spam, automated mail, duplicates; with thresholds] - Grouping: [how records were grouped into conversations, threads or documents] - Conversion: [tools and settings, e.g. Office files to Markdown, OCR engine] - Records removed in processing: [count and reasons] ## 6. De-identification | Field | Value | |---|---| | Policy version | [v1] | | Token types used | [PER, ORG, EMAIL, PHONE, ADDR, ID, URL, SECRET] | | Token scope | [numbered per conversation / per dataset] | | Date handling | [e.g. constant offset per dataset, weekday preserved] | | Categories removed outright | [e.g. health, union membership, credentials] | | Detection methods | [directory-based lists, pattern rules, NER models, human review] | | Human review sample | [number of records reviewed, by whom] | - Leak-scan results: [counts per identifier type found after processing, and what was done] - Residual risk: [what an attacker with outside knowledge could still learn, and why the supplier judges it acceptable] - Re-identification test, if run: [method and result] ## 7. Format and files | File | Description | Records | SHA-256 | |---|---|---|---| | [conversations.jsonl] | [description] | [count] | [hash] | | [index.jsonl] | [description] | [count] | [hash] | - Encoding: [UTF-8] - Field reference: [link to schema] ## 8. Quality checks | Check | Method | Result | |---|---|---| | [Turn order consistent] | [automated, full set] | [pass, 0 failures] | | [Reply links valid] | [automated, full set] | [pass] | | [Language tags correct] | [manual, 200 records] | [97% agreement] | | [Conversion fidelity] | [manual, 50 documents] | [notes] | ## 9. EU AI Act inputs For providers of general-purpose AI models. References: Article 53(1)(a), (c) and (d) of Regulation (EU) 2024/1689, Annex XI, and the Commission template for the public summary of training content. - Modality: [text / code / other] - Size range for the summary: [less than 1 billion tokens / 1 billion to 10 trillion tokens / more than 10 trillion tokens] - Source category in the summary template: [2.2.1 commercially licensed by rightsholders or their representatives / 2.2.2 private dataset from another third party] - Nature of the content: [e.g. business communications; personal data after pseudonymisation; no synthetic data] - Linguistic characteristics: [languages and registers] - Text and data mining opt-outs: [not applicable to licensed private records / describe] - Measures against illegal content: [what was screened for and how] - Data points, scope and main characteristics (Annex XI): [short statement] - How the data was obtained and selected (Annex XI): [short statement] - Methods used to detect unsuitable sources and identifiable biases (Annex XI): [short statement] ## 10. Distribution and licence - Licence type: [non-exclusive / exclusive window until YYYY-MM-DD] - Term and territory: [dates; worldwide or named regions] - Onward transfer: [allowed to affiliates, contractors, nobody] - Deletion or return at end of term: [obligation and deadline] - Model weights after termination: [what happens to models already trained] ## 11. Maintenance - Contact for questions and corrections: [role and email] - How errors are reported and fixed: [process, response time] - Withdrawal requests from individuals or rightsholders: [process] - Planned updates: [none / schedule] ## 12. Change log | Version | Date | Change | |---|---|---| | 1.0 | [YYYY-MM-DD] | First delivery |