Datasets for people who build with them.
Twelve curated datasets — geospatial, retail, climate and text — cleaned, versioned and served from one endpoint. Parquet, CSV, GeoJSON and NetCDF, with schemas that do not change under you.
Recently updated
Every release is checksummed. Older versions stay available for 90 days, so a model you trained last month still reproduces.
Built to be boring to integrate
Stable schemas, predictable file names, one checksum file per release.
Formats
Parquet with Snappy for tabular data, GeoJSON and GeoParquet for geometry, NetCDF for gridded climate series, CSV where someone downstream insists.
Licensing
CC BY 4.0 for most sets, ODbL for geospatial derivatives, CC BY-NC 4.0 for retail samples. Attribution file ships inside every archive.
Versioning
Releases are named yyyy.mm. Files are immutable; corrections ship as a new month, never as an edit in place.
A small catalogue, maintained on purpose
Deepline Data is a three-person team that cleans and publishes datasets we kept rebuilding for our own work. Nothing here is scraped live: each set is normalised once, documented, and then left alone except for scheduled refreshes.
We keep the catalogue small. If a dataset stops being useful, it is retired and its page says so, instead of quietly rotting behind a dead link.