Text-to-SQL

Schema linking and SQL generation experiments for BIRD and Spider.

Repository Layout

• schema_retriever/: table-first schema linking and SQL generation code
• schema_retriever/scripts/run_schema_linking.py: retrieve relevant tables and columns
• schema_retriever/scripts/run_sql_generation.py: generate SQL from schema-linking outputs
• schema_retriever/scripts/export_embedding_inputs.py: export embedding input text for inspection
• schema_retriever/scripts/inspect_generation_logprobs.py: inspect token-level generation confidence

Large local datasets are intentionally excluded from git. Place them under:

• datasets/bird/...
• datasets/spider/...

Generated outputs under schema_retriever/data/ are also excluded.

Quick Start

conda run -n t2s python schema_retriever/scripts/run_schema_linking.py \
  --dataset bird \
  --split dev \
  --local-files-only \
  --overwrite

conda run --no-capture-output -n t2s python schema_retriever/scripts/run_sql_generation.py \
  --dataset bird \
  --split dev \
  --overwrite

See schema_retriever/README.md for more details.