What is metadata?
Metadata is the descriptive information attached to each sample in a dataset, separate from the raw content and the task labels. For an image that might include when and where it was captured, the camera or sensor used, resolution, exposure, or the pipeline stage that produced it. It gives every sample a set of handles you can query.
Rich metadata is what turns a pile of files into a dataset you can actually navigate.
Key takeaways
- Metadata is descriptive attributes about each sample.
- It enables filtering, slicing, and sorting at scale.
- It is distinct from raw content and from task labels.
How it works
Metadata is collected at capture or generated during processing, then stored alongside each sample. Queries and filters over those fields let you assemble targeted subsets, such as all low-light images from one sensor, and define the slices used to evaluate models. It also feeds data lineage by recording how each sample was produced.
Why it matters
At scale you cannot inspect data one item at a time, so the ability to filter and group by metadata is what makes curation tractable. Good metadata underpins slicing, targeted collection, and lineage, and its absence is a common reason datasets become unmanageable.
Frequently asked questions
How is metadata different from labels?
Labels are the targets a model learns to predict, while metadata is descriptive context about the sample that is generally not the prediction target.
Why is metadata important for curation?
It lets you filter, slice, and understand large datasets efficiently, which is the basis of targeted evaluation and data collection.
Related terms