cwolff/testpile

GitTables-NG and SchemaPile-NG

Tables and database schemas from permissively licensed public GitHub repositories, snapshot of 2026-10-14.

config	records
gittables-ng	265,796 tables	239,460,508 rows stored (up to 10 000 per table)
schemapile-ng	25,447 schemas	259,111 tables

From 27,817 repositories.

from datasets import load_dataset
tables = load_dataset("cwolff/testpile", "gittables-ng", split="full")
schemas = load_dataset("cwolff/testpile", "schemapile-ng", split="full")

How it was built

• Repositories come from a uniform random sample of GitHub, license-qualified repository search and code search.
• Only permissive licenses are admitted, checked before anything is downloaded and again against the license file in the clone. Records from files near a restrictive data-license statement (non-commercial, ODbL, ...) are left out.
• Each repository is processed in a sandbox; CSV/TSV, JSON Lines, Excel/ODS, Parquet and SQLite/DuckDB files become tables, .sql/.ddl files with CREATE TABLE, SQLite/DuckDB databases and Prisma schemas become schemas (text formats also gzipped). A Prisma schema is converted to the tables Prisma Migrate creates for its database: SQL types, defaults, keys and foreign keys as in that database. SQL migration folders are skipped. Identical files are kept once.
• Within a repository: one-table SQL files of a folder are merged into one schema; older versions, the same schema for other database systems, tables whose rows are in a larger version of themselves and copies of a schema structure are left out; at most 5 tables per family (same repository, same columns) are included, the largest (182,126 more left out).
• Columns detected as personal data (emails, phone numbers, IBANs, names next to those) are replaced with Faker values; PII_REPLACED_COLUMNS lists them per table.

Fields

schemapile-ng uses the schema of SchemaPile plus CREATED_AT (drop that column to concatenate the two). IS_INDEX and INDEXES are not extracted and stay empty; VALUES (enum values), ON_DELETE and ON_UPDATE are filled for schemas from Prisma files.

gittables-ng, one record per table: ID (blob SHA, plus /<sheet or table>), FILENAME, URL (file at the commit), LICENSE, PERMISSIVE, REPO, PATH, TABLE_NAME (Excel sheet, database table or SQL INSERT table), COMMIT_SHA, BLOB_SHA, FAMILY_ID (same repository and columns), COLUMNS (header as found), ROWS (cells as strings), ROWS_TOTAL, ROWS_STORED, PII_REPLACED_COLUMNS.

CREATED_AT (both configs): when the source file was created, the author date of the first commit that touched it (for a schema merged from a folder: the folder); empty where GitHub could not date it.

Licenses and attribution

Each record keeps the license of the repository it comes from. ATTRIBUTION.csv lists source repository, commit, path, license and copyright line for every record, which is the attribution CC-BY-4.0 and similar licenses require.

license	records
MIT	155,416
Apache-2.0	65,574
CC-BY-4.0	20,505
BSD-3-Clause	16,560
BSD-2-Clause	9,708
CC0-1.0	9,070
Unlicense	3,977
MIT-0	2,218
BSL-1.0	1,882
ISC	1,822
WTFPL	1,466
PostgreSQL	1,324
0BSD	1,032
Zlib	689