TrustSQL: Benchmarking Text-to-SQL Reliability with Penalty-Based Scoring

Gyubok Lee, Woosog Chay, Seonhee Cho, Edward Choi

Keywords: Text-to-SQL, Text-to-SQL Reliability, database question-answering

Abstract: Text-to-SQL enables users to interact with databases using natural language, simplifying information retrieval. However, its widespread adoption remains limited for two main reasons: (1) existing benchmarks focus solely on feasible questions that can always be mapped to SQL queries, overlooking infeasible questions that cannot, and (2) current models lack abstention mechanisms, posing the risk of providing incorrect answers. To address these gaps, we introduce TrustSQL, a new benchmark designed to evaluate text-to-SQL reliability. At its core is the proposed Reliability Score (RS), which quantifies a model's helpfulness (correct answers) relative to its harmfulness (incorrect answers weighted by a user-defined penalty). TrustSQL is constructed by re-annotating three datasets—ATIS, Advising, and EHRSQL—while incorporating infeasible questions to enable comprehensive evaluations across diverse model inputs. We evaluate text-to-SQL models integrated with various abstention mechanisms, leveraging classification and uncertainty estimation methods. Our experiments reveal that only a few models achieve positive scores (i.e., helpfulness outweighing harmfulness) under high-penalty settings, indicating that most models are unsuitable for deployment in safety-critical scenarios. This underscores the need to develop models that not only improve SQL generation but also guarantee a certain degree of reliability, ensuring safe deployment.

Primary Area: datasets and benchmarks

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.

Reciprocal Reviewing: I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Submission Number: 10626