What is a dataset?
A dataset is a defined collection of records gathered for a specific purpose, such as training or evaluating a machine learning model. Most business datasets are tables. Each row is one example, such as a transaction or a patient visit, and each column is a feature or a label. A dataset can also be a folder of images, documents, or recordings with annotations attached. Outside machine learning, the word covers any named collection of data published for use, such as a city’s open-data release of building permits.
A production database changes constantly as the business runs. A dataset is a fixed extract with a stated scope: which source systems, which time period, and which records were excluded and why. Freezing and versioning the extract, with its data lineage recorded, lets a team rebuild a model exactly and show what it learned from.
Datasets in practice
An accounting firm wants a model that suggests general ledger codes for client invoices. Its dataset has one row per invoice line, with the vendor, description, amount, and the code a staff accountant assigned. Voided invoices are excluded, and the rows are split into training, validation, and test portions. Before pooling records across clients, the firm checks its engagement letters, because some clients restrict how their data may be used. A one-page datasheet, following the practice known as datasheets for datasets, records the source, period, exclusions, and owner for the next team and for auditors.