OptimalSolution/databank

multimod data bank

An index of curated source datasets. Each SOURCE is its own dataset repo, because a dataset card carries one license: and these sources do not share one.

source	repo	licence	contents	size
US Input-Output Accounts (BEA)	OptimalSolution/databank-bea-io	public-domain	6 datasets	1 MB
Ember	OptimalSolution/databank-ember	CC-BY-4.0	19 datasets	27 MB
FIGARO	OptimalSolution/databank-figaro	CC-BY-4.0	27 datasets	862 MB
Natural Earth	OptimalSolution/databank-basemaps	public-domain	2 layers, 4 resolutions	41 MB

Fetching, not browsing

These repos are built to be QUERIED, not downloaded. The files are parquet and Hugging Face serves HTTP range requests, so a client reads the footer, picks the row groups it needs and pulls only those column chunks. Measured on an 11.8 MB partition of a 165.9M-row table: 0.85 MB fetched, 7% of the file, for a real slice.

index.json is what makes that possible without a directory listing. It carries every source, dataset and layer with its licence, coverage, size, column list, partition key, manifest hash and the relative path of every file -- so a URL can be constructed rather than discovered. It is generated from the manifests and never hand-edited.

library(duckdb); library(jsonlite)
idx <- fromJSON(paste0(hub, "/resolve/main/index.json"),
                simplifyVector = FALSE)
# pick a source and dataset, then one entry from its `files` list
con <- dbConnect(duckdb())
dbExecute(con, "INSTALL httpfs; LOAD httpfs;")
dbGetQuery(con, sql)   # read_parquet(url) over HTTP, no download

import duckdb
duckdb.sql("INSTALL httpfs; LOAD httpfs;")
duckdb.sql(f"SELECT industry, sum(value) FROM '{url}' GROUP BY 1").df()

The dataset viewer works too, but it is a preview: splits are named train because that is Hugging Face's default for a dataset with no train/test division, not because any of this is training data.