We further evaluate DatA-SQL-1.5B on the BIRD development set using self-consistency voting.
Under Vote@8, our model achieves an execution accuracy (EX) of 55.86 %, establishing a new state of the art among open-source Text-to-SQL systems at the 1.5 B scale.
This result demonstrates that even compact models can acquire strong reasoning and SQL generation abilities when trained with our reasoning-guided data augmentation.
The lightweight design ensures low training and inference costs while maintaining high execution robustness.
Overall, DatA-SQL-1.5B achieves competitive accuracy compared with much larger GPT-based pipelines, highlighting the efficiency and practicality of our approach.
