The text-to-SQL (T2SQL) field has grown signifi- cantly with the widespread application of neural architectures and large language models (LLMs)This enables non-technical users to query databases using natural language. Unlike early rule-based and templat

The dataset contains 146 different rigorously selected and modified text to SQL tasks covering 11 domains, as well as all unresolved errors made by different GPT models in this dataset along with their error analysis. The goal of this dataset is to evaluate the LLM's reasoning ability in the case of zero-shot prompting and to provide the causes of the LLM's errors for analysis in future studies. For this reason, unlike other modern T2SQL benchmarks, this dataset does not take data directly from real life but filters out all “dirty” data, i.e., data that is logically incorrect or ambiguous, to make sure that all tasks are possible to accomplish correctly and accurately.