Lexsi DataAgent β Tiny Aya LoRA (Track 3, Statement 6)
LoRA adapter (r=16) fine-tuned on 26 verified question-to-SQL examples for a read-only business-data agent (synthetic SQLite shop database: 5 customers, 4 products, 10 orders). The adapter teaches Tiny Aya Global to emit a single read-only SELECT for six narrow question patterns (revenue by product, top customer, order counts by status, average order value, revenue in a date range, orders for a customer).
Training data: umarshareef0165/lexsi-dataagent-sft (every example's SQL was
executed against the seed database and verified before training). Evaluation:
12 held-out task questions (exact eval questions never appear in training),
base vs LoRA, scored by executing the generated SQL and checking the answer
β see lexsi-dataagent-model-eval.json in the dataset repo and the
write-up. This is a prototype-scale adapter (26 examples); it is not a
general text-to-SQL model, and that limit is stated in the write-up.