FDAbench-Full is the first heterogeneous data analysis task benchmark set for data agents, released by Nanyang Technological University, National University of Singapore and Huawei Technologies Co., Ltd. in 2025. The related paper results are "FDABench: A Benchmark for Data Agents on Analytical Queries over Heterogeneous Data", which aims to evaluate the model's capabilities in database query generation, SQL understanding, and financial data analysis.

The dataset contains 2,007 high-quality analysis tasks, covering a diverse range of data domains, difficulty levels, and task categories. Each example includes complete metadata fields, including: task_id (task unique identifier), instance_id (instance identifier), db (database name/identifier), level (difficulty level: easy/medium/hard), database_type (database system type), question_type (question category), tools_available (list of available tools), and query (main question/query text).

Dataset structure

The dataset contains three task types:

• Single-choice questions: There are 579 carefully designed questions, each with only one correct answer. They are mainly used to test the model's understanding of database concepts and SQL queries.
• Multiple-choice questions (Multiple): A total of 760 complex questions with multiple possible correct answers. They include precise numerical calculation results and conclusions based on reasoning, and are used to evaluate the model's comprehensive performance in data analysis and reasoning capabilities.
• Report Generation (report): A total of 668 questions require the generation of detailed analysis reports, testing the data agent's ability to conduct comprehensive analysis in a multi-data source environment, and providing a standard report as a comparative evaluation benchmark.