Akarshika/qwen25-3b-text2sql-lora

Qwen2.5-3B Text-to-SQL (LoRA adapter)

LoRA adapter that fine-tunes Qwen2.5-3B-Instruct for converting natural language questions into SQL queries, given the database schema in the prompt.

Exact-match accuracy: 3/100 (base) → 75/100 (fine-tuned) on a held-out test set from b-mc2/sql-create-context. Strict exact-match scoring; manual inspection shows several functionally-equivalent misses, so this is a conservative floor.

• Method: QLoRA (r=16, alpha=16), all attention + MLP projections
• Only 0.96% of parameters trained (29.9M of 3.1B)
• Trained in ~40 min on free Kaggle T4 GPUs

Full code, eval harness, and the v1→v2 iteration story: https://github.com/TripathiAkarshika/qwen-text2sql-lora