AMD QLoRA SQL Assistant fine-tunes Qwen2.5-3B-Instruct for schema-conditioned SQL generation on AMD ROCm. The public repository includes a reproducible QLoRA pipeline, SQL safety checks, a FastAPI API, a Streamlit client, and base-versus-adapter evaluation. The API accepts a question and schema separately.
On AMD MI300X, an initial 19-example smoke run showed no exact-match gain. A separate one-epoch run used 1,000 filtered SQL-Create-Context rows (800 train, 100 validation, 100 test). On that random held-out split, base versus adapter scored 0/100 vs 61/100 normalized exact match, 27/100 vs 100/100 SQL parse-valid, and 100/100 safety-pass for both. On the original five-row synthetic holdout, both scored 0/5 exact match.
An independent 30-row BIRD Mini-Dev SQLite check measured execution accuracy at 4/30 for both base and adapter; the adapter was parse-valid on 25/30 vs 16/30 for base. This does not demonstrate reliable SQL correctness. The BIRD and SQL-Create-Context results are small samples; the latter is derived from WikiSQL/Spider and may have related patterns across splits. Broader semantic and execution accuracy remain unverified. The Streamlit demo was tested over a private SSH tunnel; no public API or model weights are deployed.
.png&w=256&q=75)
.png&w=640&q=75)