Qatch-Studio: A Testbench for Text2SQL and Table Question Answering on Proprietary Data

abular Representation Learning (TRL) models and Large Language Models (LLMs) are increasingly used for Table Question Answering (TQA) and Text2SQL (T2S), yet public benchmarks fail to capture the diversity of enterprise datasets. We introduce Qatch-Studio, a flexible framework for evaluating TRL and LLM performance on SQL-centric tasks using proprietary data while preserving confidentiality. The entire pipeline, test generation, model inference, and evaluation, runs locally under user control with no data outsourcing. Qatch-Studio generates customizable test suites covering critical SQL operations (e.g., null handling, joins, aggregations) and executes them on open- or closed-source models, enabling systematic robustness assessment on user-specific data. Its extensible design allows tailoring test generation to domain-specific schemas and query types. Qatch-Studio is available online at https://huggingface.co/spaces/simone-papicchio/qatch-demo, and a demonstration video is available at https://youtu.be/1VbhOZcuZ60.