BIRD Text2SQL Bench: 9,428 Training Examples for Database-Grounded NLP

9,428 training examples convert the BIRD benchmark into OpenAI Messages format for direct use in supervised fine-tuning pipelines. This dataset adapts the official BIRD training set, which uses 70 real-world, large-scale databases across 37+ domains like e-commerce, sports, education, and healthcare. It restructures prompts to include evidence hints and provides both a consolidated messages field and separate system/user/assistant string columns.

Use Cases

• Supervised fine-tuning of LLMs for text-to-SQL based on the structured system/user/assistant dialogue format.
• Benchmark evaluation of text-to-SQL models using the pre-formatted prompts derived from the BIRD benchmark.
• Training models to incorporate external evidence or hints, based on the integrated ### Hint paragraphs in user prompts.
• Studying the performance of SQL generation on complex, real-world databases spanning multiple domains.

Strengths

• Contains 9,428 structured training examples formatted for direct SFT pipeline input.
• Based on the BIRD benchmark, which uses 70 real-world, large-scale databases.
• Covers a wide range of 37+ application domains including e-commerce, sports, education, and healthcare.
• Provides both a consolidated OpenAI Messages format and separate, decomposed string columns for flexibility.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row count is known (9,428), but other scale details like file formats and total size are unknown.
• The description is specific, but the original dataset's license and author organization are not provided.

Provenance

Source

lianghsun on Hugging Face, adapting the official BIRD (BIg Bench for Large-Scale Database Grounded Text-to-SQL) training set.

Collection Method

Conversion of the BIRD benchmark training data into OpenAI Messages format, with evidence integrated into prompts.

Freshness

Last updated 2026-04-15 03:13:03; freshness should be verified.