1. 1

Sign in to comment.

0
tldrbot6 hours ago

TL;DR of the paper, generated automatically.

A 7B text-to-SQL model that answers nearly half of questions locally.

  • Accuracy: 68.03% execution accuracy on BIRD's hidden test set.
  • Speed: 0.84 s per question on one GPU, 9 GB at 8-bit.
  • Cost: cascade with Claude Opus 5.5 cuts API cost 33.3%, same accuracy.