SPENCE: A Syntactic Probe for Detecting Contamination in NL2SQL Benchmarks

TL;DR

This paper introduces SPENCE, a framework for detecting contamination in NL2SQL benchmarks by analyzing how models' accuracy changes with syntactic variations of test queries.

Contribution

The paper presents a novel syntactic probing method to quantify contamination in NL2SQL benchmarks and evaluates its effectiveness across multiple datasets and models.

Findings

Older benchmarks like Spider show high contamination levels.

Recent datasets like BIRD are largely uncontaminated.

Syntactic robustness varies with benchmark age and design.

Abstract

Large language models (LLMs) have achieved strong performance on natural language to SQL (NL2SQL) benchmarks, yet their reported accuracy may be inflated by contamination from benchmark queries or structurally similar patterns seen during training. We introduce SPENCE (Syntactic Probing and Evaluation of NL2SQL Contamination Effects), a controlled syntactic probing framework for detecting and quantifying such contamination. SPENCE systematically generates syntactic variants of test queries for four widely used NL2SQL datasets-Spider, SParC, CoSQL, and the newer BIRD benchmark. We use SPENCE to evaluate multiple high-capacity LLMs under execution-based scoring. For each model, we measure changes in execution accuracy across increasing levels of syntactic divergence and quantify rank sensitivity using Kendall's tau with bootstrap confidence intervals. By aligning these robustness trends with benchmark release dates, we observe a clear temporal gradient: older benchmarks such as Spider exhibit the strongest negative values and thus the highest likelihood of training leakage, whereas the more recent BIRD dataset shows minimal sensitivity and appears largely uncontaminated. Together, these findings highlight the importance of temporally contextualized, syntactic-probing evaluation for trustworthy NL2SQL benchmarking.