Putting Task Expertise into RL Achieves SOTA Performance on Text-to-SQL
Comments
Ranked by a public rubric, not by votes: what a comment adds counts, not who wrote it. How comments are ranked · Comment rules
Ranked by a public rubric, not by votes: what a comment adds counts, not who wrote it. How comments are ranked · Comment rules
Summary
A research team (UIUC Kang Lab) reports ReViSQL-K2.6, a single fine-tuned model trained with reinforcement learning on a cleaned training set, BIRD-Platinum, that reaches 88.55% on the expert-verified Arcwise-Plat-SQL benchmark and exceeds the 92.96% human proxy with 16-sample self-consistency.