gbag-bench
收藏资源简介:
GBAG-Bench(Grounded BI Answer Generation)是首个公开基准数据集,用于衡量大型语言模型(LLM)将SQL查询结果忠实、完整、有洞察力地解释为自然语言答案的能力。该数据集旨在弥补现有基准(如Spider、BIRD、WikiSQL)仅关注SQL生成正确性而忽略自然语言答案生成质量的不足。数据集包含35个精心标注的英文问题,覆盖三个公共示例数据库:Sakila(DVD租赁,约5MB)、Chinook(音乐商店,约1MB)和Northwind(批发贸易,约25MB)。每个样本以JSON对象形式存储,包括唯一标识符(id)、数据库名称(database)、自然语言问题(question)、参考SQL查询(gold_sql)、参考自然语言答案(gold_answer)、难度等级(difficulty,1-10级)、问题类别(category)和预期洞察列表(expected_insights)。核心任务是在给定自然语言问题、SQL查询结果和完整结果集的情况下,生成忠实于数据、完整且具洞察力的自然语言答案。评估采用LLM作为评判者,基于忠实性、完整性和洞察力进行综合打分(权重分别为0.50、0.30和0.20)。数据集适用于表格问答、文本生成等任务,规模较小(<1K),主要用于基准测试和模型能力评估。
GBAG-Bench (Grounded BI Answer Generation) is the first public benchmark dataset designed to measure the ability of large language models (LLMs) to faithfully, completely, and insightfully interpret SQL query results into natural language answers. It addresses the limitation of existing benchmarks (such as Spider, BIRD, and WikiSQL) that focus only on SQL generation correctness while neglecting the quality of subsequent natural language answer generation. The dataset contains 35 carefully annotated English questions covering three well-known public example databases: Sakila (DVD rental, approximately 5MB), Chinook (music store, approximately 1MB), and Northwind (wholesale trade, approximately 25MB). Each sample is stored as a JSON object with fields including unique identifier (id), database name (database), natural language question (question), reference SQL query (gold_sql), reference natural language answer (gold_answer), difficulty level (difficulty, 1-10), question category (category), and a list of expected insights (expected_insights). The core task is to generate a natural language answer that is faithful to the data, complete in covering key information, and insightful, given a natural language question, the executed SQL query (gold reference), and the full result set (rows and columns). Evaluation uses an LLM as a judge, scoring based on handcrafted gold answers and expected insight lists, with a composite score formula: GBAG score = 0.50 × faithfulness + 0.30 × completeness + 0.20 × insightfulness, where faithfulness has the highest weight, reflecting the severity of hallucinated numbers leading to incorrect real-world decisions in BI scenarios. The dataset is suitable for tasks such as table question answering, text-to-text generation, and text generation, particularly for evaluating LLMs in grounded generation, reducing hallucinations, data analysis, and business intelligence report generation. The dataset is small in scale (<1K) and primarily used for benchmarking and model capability assessment rather than drawing final conclusions.





