Dataset Datasheet Generator

The seven datasheet sections as seven boxes, producing a document and a JSON file you can check in beside the data.

Dataset name

Version

RowsHowever you count them. Say what a row is below.

Covering what period

LicenceOn the data, which is not always the licence on the code.

Maintainer contact

Where it livesA path, a bucket, a table — wherever the next person has to look.

MotivationWhy was this dataset created, and for what task? Who created it, and who paid for it?

CompositionWhat does one row represent? How many rows are there, and is that a sample of something larger? Does it contain data about people? Anything confidential? Are there errors, duplicates or known noise in it?

Collection processHow was the data acquired — observed, reported, scraped, generated? Over what period, and by whom? If it is about people, were they told, and did they agree?

Preprocessing, cleaning and labellingWhat was removed, normalised, deduplicated or filtered? Who labelled it, under what instructions, and how was agreement measured? Is the raw data still available?

UsesWhat has it already been used for? What should it NOT be used for, and why? What in the collection or cleaning could bias a downstream model?

DistributionWho can have it, and under what licence? Are there third-party terms or export restrictions on it?

MaintenanceWho maintains it, and how are corrections submitted? Will it be updated, and how are older versions kept available?

Output format

Print the questions above each answer, so a reader can see what was asked

Sections answered

3 of 7

Every unanswered section is printed into the document with its questions, so the gap is visible to whoever inherits the dataset instead of being invisible to everyone.

Sections answered: 3
Sections left blank: 4
Questions in the framework: 19
Words: 319

Your datasheetRead-only. Select all and copy.

# Datasheet — support-tickets-2026

- **Dataset:** support-tickets-2026
- **Version:** 1.0
- **Rows:** 48,000

## Motivation

> Why was this dataset created, and for what task?
> Who created it, and who paid for it?

Built to train and evaluate a router that sends an inbound support message to one of eleven queues. Created by the support tooling team.

## Composition

> What does one row represent?
> How many rows are there, and is that a sample of something larger?
> Does it contain data about people? Anything confidential?
> Are there errors, duplicates or known noise in it?

One row is one inbound message plus the queue a human eventually moved it to. Everything that arrived through the web form or by email is included; live chat is not. Around 3% of rows are duplicates, where a customer sent the same message twice.

## Collection process

> How was the data acquired — observed, reported, scraped, generated?
> Over what period, and by whom?
> If it is about people, were they told, and did they agree?

Exported from the ticketing system in one query covering the period above. Nobody was asked separately: the messages are the customers' own support requests, retained under the existing support-data retention policy.

## Preprocessing, cleaning and labelling

> **Not answered.** What was removed, normalised, deduplicated or filtered? Who labelled it, under what instructions, and how was agreement measured? Is the raw data still available?

## Uses

> **Not answered.** What has it already been used for? What should it NOT be used for, and why? What in the collection or cleaning could bias a downstream model?

## Distribution

> **Not answered.** Who can have it, and under what licence? Are there third-party terms or export restrictions on it?

## Maintenance

> **Not answered.** Who maintains it, and how are corrections submitted? Will it be updated, and how are older versions kept available?

The same thing as JSONFor checking a datasheet into the repository next to the data, where a script can tell whether it is complete.

{
  "dataset": "support-tickets-2026",
  "version": "1.0",
  "rows": "48,000",
  "covering": "",
  "licence": "",
  "contact": "",
  "location": "",
  "sections": {
    "motivation": {
      "questions": [
        "Why was this dataset created, and for what task?",
        "Who created it, and who paid for it?"
      ],
      "answer": "Built to train and evaluate a router that sends an inbound support message to one of eleven queues. Created by the support tooling team."
    },
    "composition": {
      "questions": [
        "What does one row represent?",
        "How many rows are there, and is that a sample of something larger?",
        "Does it contain data about people? Anything confidential?",
        "Are there errors, duplicates or known noise in it?"
      ],
      "answer": "One row is one inbound message plus the queue a human eventually moved it to. Everything that arrived through the web form or by email is included; live chat is not. Around 3% of rows are duplicates, where a customer sent the same message twice."
    },
    "collection": {
      "questions": [
        "How was the data acquired — observed, reported, scraped, generated?",
        "Over what period, and by whom?",
        "If it is about people, were they told, and did they agree?"
      ],
      "answer": "Exported from the ticketing system in one query covering the period above. Nobody was asked separately: the messages are the customers' own support requests, retained under the existing support-data retention policy."
    },
    "preprocessing": {
      "questions": [
        "What was removed, normalised, deduplicated or filtered?",
        "Who labelled it, under what instructions, and how was agreement measured?",
        "Is the raw data still available?"
      ],
      "answer": null
    },
    "uses": {
      "questions": [
        "What has it already been used for?",
        "What should it NOT be used for, and why?",
        "What in the collection or cleaning could bias a downstream model?"
      ],
      "answer": null
    },
    "distribution": {
      "questions": [
        "Who can have it, and under what licence?",
        "Are there third-party terms or export restrictions on it?"
      ],
      "answer": null
    },
    "maintenance": {
      "questions": [
        "Who maintains it, and how are corrections submitted?",
        "Will it be updated, and how are older versions kept available?"
      ],
      "answer": null
    }
  }
}

Sections still unanswered — they appear as holes in the document:

Covering
Licence
Maintainer contact
Where it lives
Preprocessing, cleaning and labelling
Uses
Distribution
Maintenance

What this assumes: the seven groups and their questions come from the datasheets-for-datasets framework; the answers are entirely yours, and nothing is inferred from one answer to fill another. The text this page opens with is an example about an invented dataset, there to show the shape of a real answer — replace it. A blank section is printed as a blank with its questions attached rather than quietly dropped, because a datasheet whose gaps are invisible is a datasheet that gets waved through. The row count and the period are text, not measurements — this page cannot see your data. Everything on this page runs in your browser. Nothing you paste is uploaded, logged or sent anywhere.

Every dataset has a person who knows what is wrong with it, and that person leaves. The questions above exist because the knowledge that disappears with them is specific and predictable: which period the scrape actually covers, why the January rows are duplicated, which label was applied by one contractor with a different instruction sheet, and the fact that the “random sample” was the first 50,000 rows by id.

The section that pays for itself

Distribution. A model trained on data you were not licensed to train on is not a data-quality problem, it is a problem you cannot fix by retraining, because the question arrives months later and the answer has to be reconstructed from memory. Writing down where each part came from and under what terms takes ten minutes now and is unbuyable later.

Answer the ones you cannot answer

“We do not know how these were labelled” is a complete and useful answer. It tells the next person not to trust the labels more than the data supports, which is the entire point of the document, and it is far better than a blank that reads as “nothing to see here”. The completeness figure above counts answers, not reassurances — write the awkward ones down and the number goes up for the right reason.

Dataset Datasheet Generator

The section that pays for itself

Answer the ones you cannot answer

The reasoning behind this

Related tools