| --- |
| license: other |
| base_model: Qwen/Qwen2.5-Coder-7B-Instruct |
| tags: [text-to-sql, bird, spider, qlora, aegis-sql] |
| --- |
| # Qwen25-Coder7B-Aegis-Text2SQL |
|
|
| QLoRA fine-tune of `Qwen/Qwen2.5-Coder-7B-Instruct` for Text-to-SQL, trained as the local-path |
| generator of the AEGIS-SQL hybrid system (three-axis constrained NL2SQL: |
| accuracy / cost / privacy). |
|
|
| - **Data**: 1,000 curated examples — 500 BIRD-train + 500 diversity-stratified |
| Spider-train (set ops, HAVING, nested, multi-join rebalanced); splits 798/101/101. |
| - **Recipe**: 4-bit NF4 QLoRA, r=32 alpha=64, lr=0.0001, |
| 3 epochs, completion-only loss, max_seq_len=4096, seed=42. |
| - **Prompt format**: chat messages; user = `Database:` + schema (Tables + Foreign |
| Keys) + `Question:` + `Return only SQL.`; assistant = bare SQL. |
| - **Quick proxy on held-out test**: see `quick_eval.json` in this repo. |
| Full BIRD EX requires the official evaluator + databases. |
|
|
| Trained with the text2sql_aegis_experiment package (notebooks/qwen_aegis.ipynb). |
| |
