Corpus-adaptive fine-tuning improves SQL schema retrieval recall

The authors recast five text-to-SQL datasets (Spider, BIRD, BEAVER, and two LiveSQLBench variants) as table and column retrieval tasks to evaluate schema linking. They propose corpus-adaptive fine-tuning for a 305M-parameter embedder, which raises average recall@10 from 60.4 to 75.6.

• The method synthesizes natural-language queries directly from the target schema corpus and mines granularity-aware hard negatives.
• It outperforms off-the-shelf text and code embedders, matching state-of-the-art results on the benchmark.
• Applying the same recipe to an 8B embedder improved its recall@10 from 77.8 to 78.4.
• Experiments confirm gains reflect transferable ability rather than data memorization.

This establishes schema linking as a standalone retrieval task and demonstrates lightweight, label-free corpus adaptation as a practical route for enterprise-scale deployment.