Datasets:
Missing test set.
Why is test set missing from the dataset ? It would have been great to include the test set, so that the community can build better model and beat the benchmark.
Thanks for your interest in our work! This dataset is created for text-to-SQL instruction tuning and our goal is to teach FM to understand text-to-SQL instruction. If you want to benchmark this dataset, you can create a test split with different strategies. You are very welcome to provide feedback/suggestions/comments.
Hi,
I am referring to reproducing the test scores mentioned in the blog post. https://www.numbersstation.ai/post/introducing-nsql-open-source-sql-copilot-foundation-models
The test set from original authors(company) would help in consistent benchmark than the community creating a separate one.
Thanks for your clarification!
We've followed the evaluation from the original benchmark (Spider: https://github.com/taoyds/spider, and GeoQuery: https://github.com/jkkummerfeld/text2sql-data/tree/master/data). And we use the Rajkumar prompt format described here (https://arxiv.org/abs/2204.00498) and you can example code here (https://github.com/NumbersStationAI/NSQL/tree/main/examples).
Hope this helps!