Datasheet for Datasets Template (Gebru et al. 2021)

Dataset motivation, composition, collection, labelling, uses, distribution, maintenance, privacy and consent

AI・データフルDatasheets for Datasets (Gebru et al., 2021)datasheet for datasets更新 2026/10/20

使い方: Use when creating or publishing a dataset for training or evaluation, ideally while it is being collected. Answer motivation and composition first, with reasons for any N/A. The main pitfall is skipping the people-related questions on consent and sensitive data.

プレビュー

Datasheet: [Dataset name] [version]

Complete datasheet for a dataset used to train, evaluate or benchmark machine learning systems. It answers, in order, the questions a careful consumer of the data would ask: why it exists, what is in it, how it was collected and processed, what it should and should not be used for, how it is distributed and who maintains it. The section structure follows Gebru et al. (2021), Datasheets for Datasets. Replace every [bracketed] value and every answer marked Example.

Document control

Field	Value
Document ID	[DS-DATASET-001]
Dataset name	[Dataset name]
Dataset version	[2.0.0]
Datasheet version	[0.1]
Status	[Draft / In review / Approved / Published / Deprecated]
Owner	[@data-owner]
Classification	[Internal / Confidential / Public]
Last updated	YYYY-MM-DD
Persistent identifier	[DOI or registry URI], checksum `sha256:[hash]`
Related documents	[Model cards using this dataset], [DPIA], [Data contract]

Approvals

Role	Name	Decision	Date
Dataset owner	[@name]	[Approved / Changes requested]	YYYY-MM-DD
Privacy officer	[@name]	[Decision]	YYYY-MM-DD
Responsible AI reviewer	[@name]	[Decision]	YYYY-MM-DD

Revision history

Version	Date	Author	Change summary
0.1	YYYY-MM-DD	[@name]	First draft during collection
1.0	YYYY-MM-DD	[@name]	Example: Published with dataset 2.0.0
1.1	YYYY-MM-DD	[@name]	Example: Added removal of 312 records after deletion requests

Motivation
Composition
Collection process
Preprocessing, cleaning and labelling
Uses
Distribution
Maintenance
Privacy, consent and data protection summary
References and glossary

1. Motivation

1.1 Purpose

Why was the dataset created, and was there a specific task or gap in mind?

[Answer. Example: "To train and evaluate a classifier that detects urgent customer support tickets. No public dataset covered support tickets with urgency labels in our product domain."]

1.2 Creators

Who created the dataset, and on behalf of which organisation?

[Answer. Example: "The Support Analytics team (four engineers, two annotation leads) at [Organisation]."]

1.3 Funding

Who funded the creation of the dataset? Name the grant or budget if any.

[Answer. Example: "Internal budget of the Customer Operations division, project code CO-2026-07."]

1.4 Other comments

[Any other context on motivation, for example a prior dataset this one replaces and why.]

2. Composition

2.1 Instances

What does each instance represent? Are there several types of instance?

[Answer. Example: "Each instance is one support ticket: subject, body, channel, product area, creation time and an urgency label. There is one instance type."]

2.2 Size and sampling

Question	Answer
How many instances are there in total, per type?	Example: 334 000 tickets; 30 000 urgent
Is it all possible instances or a sample?	Example: A sample: all English tickets from 18 months, excluding spam
If a sample, how was it drawn and is it representative?	Example: Complete within the period; under-represents chat, which launched late in the period

2.3 Data fields

Field	Type	Description	Missing values	Example
`ticket_id`	string	Pseudonymous identifier	None	`T-8f31c2`
`subject`	string	Ticket subject, PII masked	2% empty	`Cannot log in after update`
`body`	string	Ticket body, PII masked, truncated at 4 000 characters	None	`Since this morning ...`
`channel`	enum	email, web, chat	None	`web`
`label_urgent`	boolean	Consensus urgency label	None	`true`
[field]	[type]	[Description]	[Rate]	[Example]

2.4 Labels and targets

Is there a label or target for each instance? What is it and how reliable is it?

[Answer. Example: "Binary urgency label from two annotators with lead adjudication; agreement before adjudication was kappa 0.81."]

2.5 Missing information, relationships and splits

Question	Answer
Is any information missing from instances, and why?	[Answer]
Are relationships between instances explicit (for example threads, same customer)?	Example: `thread_id` links follow-up tickets; customer identity removed
Are there recommended splits?	Example: Time-based: train to YYYY-MM, validation next 2 months, test last 3 months
Are there errors, noise or redundancies?	Example: About 4% near-duplicate tickets during incidents
Is the dataset self-contained or does it link to external resources?	Example: Self-contained; attachment links removed

2.6 Confidential and sensitive content

Question	Answer
Could any data be confidential (legal privilege, contracts, non-public communications)?	[Answer]
Could any data be offensive, insulting, threatening or cause anxiety if viewed?	Example: Yes, about 1% of tickets contain abusive language
Does the dataset relate to people?	Example: Yes, customers who wrote tickets
Does it identify subpopulations (age, gender, region)? How?	Example: No explicit attributes; language can be inferred
Can individuals be identified, directly or in combination with other data?	Example: Possible from free text despite masking; treated as personal data
Does it contain special category data (health, ethnicity, religion, sexuality, biometrics, criminal records, financial data)?	Example: Incidental mentions of health in about 0.3% of tickets

2.7 Coverage summary

Summarise how the dataset is distributed over the characteristics most likely to affect models trained on it. Consumers use this table to decide which groups or conditions they must evaluate separately.

Characteristic	Category	Share of instances	Share in target population	Gap
Channel	Example: Chat	6%	21% (current traffic)	Under-represented
Language background	Example: Second-language writers (estimated)	14%	18%	Slightly under
Period	Example: Incident days	9%	7%	Slightly over
[Characteristic]	[Category]	[%]	[%]	[Gap]

3. Collection process

3.1 Acquisition

Question	Answer
How was the data acquired: directly observed, reported by subjects, inferred or derived?	Example: Directly observed: exported from the ticketing system
What mechanisms or procedures were used (hardware, software API, manual curation)?	Example: Nightly export via the ticketing API, validated by row counts
If sampled, what was the sampling strategy?	[Answer]
Who was involved in collection and how were they compensated?	Example: Employees as part of normal duties
Over what timeframe was the data collected? Does it match the creation time of the content?	Example: YYYY-MM to YYYY-MM, same as ticket creation

3.2 Ethical review

Was any ethical review process conducted? Give the outcome and a link if possible.

[Answer. Example: "Reviewed by the Responsible AI board on YYYY-MM-DD, approved with conditions: PII masking and a 24-month retention limit."]

3.3 Data about people: notice and consent

Question	Answer
Was the data collected from the individuals directly or from third parties?	[Answer]
Were individuals notified about the collection? How?	Example: Privacy notice section 4 states tickets may be used to improve support tools
Did individuals consent? How was consent requested and recorded?	Example: Not consent-based; legal basis is legitimate interest, with an assessment on file
Is there a way to revoke consent or object?	Example: Deletion and objection requests remove the ticket at the next release
Has a data protection impact assessment been done?	Example: Yes, DPIA-2026-11

4. Preprocessing, cleaning and labelling

4.1 Processing steps

Step	What was done	Tool and version	Records affected
1	Example: Removed spam and internal test tickets	Rules in `filters.yml` v3	12 400 removed
2	Example: Masked emails, phone numbers, names	NER model [name, version] plus regex	All
3	Example: Truncated body at 4 000 characters	Pipeline v2.1	1.8%
4	Example: Deduplicated exact copies within 1 hour	Hash on normalised text	3 100 removed
[n]	[Step]	[Tool]	[Count]

4.2 Labelling

[Annotator pool, guidelines version, training, quality control, agreement statistics and how disagreements were resolved. Example: "Twelve trained agents labelled with guideline v4; 10% double-labelled; kappa 0.81; lead adjudication for all disagreements."]

4.3 Raw data and software

Question	Answer
Was the raw data saved? Where, and who can access it?	[Answer]
Is the preprocessing software available?	Example: Yes, repository `data-pipelines/tickets`, tag `ds-2.0.0`

5. Uses

5.1 Uses so far and repository of uses

[Answer. Example: "Training and evaluating urgency model 1.x (see model card MC-URGENCY-001). A list of uses is kept in the dataset registry."]

5.2 Other suitable tasks

[Answer. Example: "Topic classification of support tickets; routing to product areas."]

5.3 Risks from composition and collection

Could anything about the composition or collection affect future uses or cause unfair treatment, legal or financial harm?

[Answer. Example: "Chat is under-represented, so models trained on this data may underperform on chat. Users should evaluate per channel."]

5.4 Uses to avoid

ID	Use to avoid	Reason
AV-01	Example: Profiling individual customers	Data collected for support improvement only
AV-02	Example: Evaluating agent performance	Labels reflect ticket content, not agent work
[AV-nn]	[Use]	[Reason]

6. Distribution

Question	Answer
Will the dataset be distributed outside the creating organisation?	Example: No, internal teams only
How is it distributed (download, API, registry)? Does it have a DOI?	Example: Internal dataset registry, versioned
When was or will it be distributed?	YYYY-MM-DD
Under what licence or terms of use?	[Licence or internal policy]
Have third parties imposed restrictions on the data?	[Answer]
Do export controls or other regulatory restrictions apply?	[Answer]

7. Maintenance

Question	Answer
Who supports and maintains the dataset?	[@data-owner], [team]
How can the owner be contacted?	[Channel, email alias]
Is there an erratum?	[Link to errata log]
Will the dataset be updated? How often, by whom, and how will consumers learn about it?	Example: Quarterly release; announced in the registry change feed
Are there retention limits for data about people? Are they enforced?	Example: 24 months from ticket creation; records past the limit dropped at each release
Will older versions remain available?	Example: Last two versions kept for reproducibility, then deleted
Can others extend or contribute? How are contributions validated?	[Answer]

7.1 Errata and change log

Record every correction and every removal of records, including removals made in response to deletion requests, so that consumers can tell whether a result was produced on affected data.

ID	Date	Version affected	Issue	Action	Fixed in
ER-01	YYYY-MM-DD	Example: 2.0.0	140 tickets labelled urgent by a misconfigured import	Labels corrected after re-review	2.0.1
ER-02	YYYY-MM-DD	Example: 2.0.1	312 records subject to deletion requests	Records removed	2.1.0
[ER-nn]	YYYY-MM-DD	[Version]	[Issue]	[Action]	[Version]

8. Privacy, consent and data protection summary

Item	Status	Evidence
Legal basis recorded	[Yes / No]	[Reference]
Notice given to data subjects	[Yes / No]	[Privacy notice section]
Data minimisation applied	[Yes / No]	[Fields dropped, masking]
Special category data handled	[Yes / No / N/A]	[Control]
Deletion and objection requests honoured	[Yes / No]	[Process]
Retention limit enforced	[Yes / No]	[Job or control]
DPIA completed	[Yes / No / Not required, reason]	[Reference]

This datasheet documents facts about the dataset. It is not legal advice; confirm data protection obligations with your privacy or legal function.

9. References and glossary

9.1 References

Reference	Use
Gebru, T., Morgenstern, J., Vecchione, B., Wortman Vaughan, J., Wallach, H., Daumé III, H., Crawford, K. (2021). Datasheets for Datasets. Communications of the ACM 64(12), 86-92.	Question structure
[Model card]	Downstream use of this dataset
[Annotation guideline vN]	Labelling rules

9.2 Glossary

Term	Meaning
Instance	One record of the dataset, such as one ticket or one image
Adjudication	Resolution of annotator disagreement by a senior reviewer
Kappa	Agreement statistic between annotators corrected for chance
Pseudonymous identifier	An ID that replaces a direct identifier but can be linked back with extra data

Standard and licence note

Structure aligned with the question sections of Gebru et al. (2021), Datasheets for Datasets (motivation, composition, collection process, preprocessing, cleaning and labelling, uses, distribution, maintenance). The questions here are paraphrased and extended in original text written for this collection and reproduce no text from the paper; read the paper for the authors' full question set. Template licence: MIT.