text2sql-data

Why we included this project

Anyone training a system that turns natural-language questions into SQL will appreciate having the standard benchmarks in one place. The repo holds cleaned, canonicalised versions of familiar corpora like ATIS, Geography, Restaurants, Scholar, IMDB, Yelp, Spider, and WikiSQL, plus a new Advising dataset, each with annotated sentences, SQL queries, a schema, and a database. Because every source is normalised into the same format, you can compare models across domains without re-normalising each one yourself. The evaluation methodology from the accompanying ACL 2018 paper is documented here too, which helps when you are designing or reporting text-to-SQL experiments. It is a convenient, citable bundle of the data you would otherwise have to assemble by hand.