This directory contains the datasets, one of which, advising, is new, while the others are modified forms of data from prior work. Our changes included:


For each dataset we provide:

Some of the databases are not included in the repository either for size or licensing reasons. They can be found as follows:

Dataset Database
Academic (MAS), IMDB, Yelp

For more information about the sources of data see the file.

Evaluation data split definition

Question split - Where possible, we follow the divisions from prior work.

Query split - Random assignment (note that this did not take into consideration the number of questions for a given query).

For the smaller datasets, we use cross-validation (randomly assigned) and provide our split definitions to enable exact replication.

For the larger datasets, at test time we train on both the training and development sets.


Each json file contains a list of queries with the following fields:

Symbol Type Meaning
query-split string Whether this is a training, development or test query [large datasets] or the split number for 10-fold cross validation [small datasets]
sentences list of mappings -
sentences/question-split string Whether this is a training, development or test question [large datasets] or the split number for 10-fold cross validation [small datasets]
sentences/text string The text of the question, with variable names
sentences/variables mapping Mapping from variable names to values
sql list of strings SQL queries with variable names. Note - we only use the first query, but retain the variants for completeness (e.g. using joins vs. conditions).
variables list of mappings -
variables/location string Whether this occurs in the SQL only, the question only, or both
variables/example string An example value that could fill the variable (in the SQL only case, this is what is used)
variables/name string The variable name
variables/type string Dataset specific type

For WikiSQL and Spider we have a few additional fields:

Symbol Type Meaning
sentences/original string The question from the original dataset, before our simple tokenisation
sentences/database string The name of the database this question is for [Spider]
sentences/table-id string The name of the table this question is for [WikiSQL]
sql-original list of strings The query from the original dataset, before our canonicalisation

Also, there are a few caveats:


    "query-split": "test",
    "sentences": [
            "question-split": "train",
            "text": "Is course number0 available to undergrads ?",
            "variables": {
                "department0": "",
                "number0": "519"
    "sql": [
    "variables": [
            "example": "EECS",
            "location": "sql-only",
            "name": "department0",
            "type": "department"
            "example": "595",
            "location": "both",
            "name": "number0",
            "type": "number"

The schema is formatted as a series of lines, each describing one field from a table:

When a value is not set (e.g. default) a - is used.

Other content

The two directories contain relevant data that we did not use in our paper: