Datasheets For Datasets: Definition and Examples

Datasheets For Datasets are a standardized documentation framework introduced by Timnit Gebru and co-authors in 2018. The concept is directly inspired by datasheets used in the electronics industry, where each component is accompanied by a document describing its characteristics, usage conditions, and limitations. Applied to datasets, this principle aims to address a glaring lack of documentation in the field of machine learning.

A datasheet for a dataset covers several essential dimensions: the motivation behind creating the dataset, the data collection process, their composition and structure, the preprocessing steps applied, recommended and discouraged uses, as well as ethical considerations related to their use. Each section is guided by specific questions that dataset creators must answer.

This approach addresses major challenges of modern AI. Without adequate documentation, practitioners risk using biased, unrepresentative, or inappropriate data for their use case, which can lead to discriminatory or unreliable models. Datasheets allow users to make informed decisions about the suitability of a dataset for their specific application.

In the context of prompt engineering, understanding datasheets is crucial because the quality of a language model's responses directly depends on the data it was trained on. Knowing the limitations and potential biases of training data helps formulate more precise prompts and interpret results critically.