Integrity audit: duplicates, degenerate records, and split overlap
I ran a reproducible integrity audit over Spider and wanted to share the results here, since they reflect well on the dataset and the method may be useful to others.
Within-split duplication
| Split | Records | Schema issues | Degenerate | Genuinely duplicated |
|---|---|---|---|---|
| train | 7,000 | 0 | 0 | 8 clusters / 16 records (0.23%) |
| dev | 1,034 | 0 | 0 | 0 |
One caveat that turned out to matter. A plain text-similarity pass flags 77 duplicate clusters in the training split. Only 8 are real. The other 69 reuse a question phrasing against a different database or with different reference SQL — deliberate construction, not redundancy. "How many players are there?" appears twice, once about a tennis dataset and once about a riding club: identical text, unrelated problems.
Anyone deduplicating Spider with an off-the-shelf similarity check will over-remove by roughly 9x.
Train/dev overlap
| Check | Matches | % of dev |
|---|---|---|
| Identical question | 6 | 0.58% |
| Identical reference SQL | 8 | 0.77% |
| Both identical | 2 | 0.19% |
| Shared db_id values | 0 | — |
The splits share no databases at all, so an identical question across them is being asked of entirely different data. Every collision is a degenerate row count landing on a coincidentally same-named table — "What is the total number of airlines?" appears against both flight_2 and flight_4, each reducing to SELECT count(*) FROM AIRLINES.
Two dev items out of 1,034 are answerable from memorized training data, both trivial counts. The split separation holds.
Reproducing
Code and full reports: https://github.com/ashishsinha1602/dataset-integrity-audit — pure standard library, every number reproduces with two commands.
There is also a hosted version that audits any dataset id: https://huggingface.co/spaces/Ashsinha1/dataset-integrity-auditor
Happy to run any variant of this if it would be useful.