| library_name: lightgbm | |
| tags: | |
| - text-to-sql | |
| - sql | |
| - reranker | |
| - lightgbm | |
| - lambdarank | |
| - duckdb | |
| - codet5 | |
| # Text-to-SQL LightGBM Reranker | |
| This repository contains a trained LightGBM LambdaRank reranker for selecting the best SQL candidate from CodeT5-generated SQL candidates. | |
| ## Performance Summary | |
| - Rows: 4701 | |
| - Candidate upper bound: 0.4482025101042331 | |
| - Rule reranker exact: 0.24803233354605403 | |
| - Validation ML exact: 0.3995749202975558 | |
| - Validation ML correct: 376 / 941 | |
| - Full ML exact: 0.43671559242714314 | |
| ## Files | |
| - sql_reranker_lightgbm.txt: trained LightGBM model | |
| - feature_columns.json: exact feature column order used during training | |
| - feature_importance.csv: LightGBM feature importance | |
| - summary.json: training summary | |
| - inference.py: helper functions for loading and scoring | |
| ## Intended Pipeline | |
| Question -> NER + value index -> CodeT5 candidates -> feature extraction -> LightGBM reranker -> best SQL | |
| ## Important | |
| This reranker does not generate SQL by itself. It only ranks already-generated SQL candidates. | |
| At inference time, use the same feature extraction logic as training and reindex with feature_columns.json. |