bevaya/FinQA

FinQA是一个用于金融数据数值推理的数据集，完整复现了FinQA数据集（Chen等人，EMNLP 2021）。该数据集包含基于标普500公司（1999-2019年）收益报告的问题，数据来源于FinTabNet数据集。每个示例配对一个金融表格和周围文本（包括表格前后的文本段落），以及一个问题、一个人类可读答案和一个结构化推理程序，该程序指定了推导答案所需的算术操作。数据集提供了详细的字段，如id、文件名、表格数据（包括原始和规范化版本）、问题、答案、执行答案、解释、标注证据行、步骤结构（表示算术操作）、程序（扁平化和嵌套形式）以及黄金证据索引。数据集分为训练集（6,251个示例）、验证集（883个示例）和测试集（1,147个示例），并排除了私有排行榜测试集。该版本保留了完整的推理注释（如program、program_re和steps），这些是FinQA的主要贡献。数据集仅包含文本，源PDF页面图像可在FinTabNet中获取，并通过id和filename字段与FinTabNet文件结构关联。许可证方面，QA注释使用MIT许可证，底层表格数据基于CDLA-Permissive-1.0许可证。FinQA is a dataset for numerical reasoning over financial data, which fully replicates the FinQA dataset (Chen et al., EMNLP 2021). This dataset contains questions based on the earnings reports of S&P 500 companies (1999–2019), with data sourced from the FinTabNet dataset. Each example pairs a financial table with its surrounding context text (including text paragraphs before and after the table), alongside a question, a human-readable answer, and a structured reasoning program that specifies the arithmetic operations required to derive the answer. The dataset includes detailed fields such as id, filename, table data (both raw and normalized versions), question, answer, executed answer, explanation, annotated evidence rows, step structure (representing arithmetic operations), programs (both flattened and nested forms), and gold evidence indices. The dataset is split into training set (6,251 examples), validation set (883 examples), and test set (1,147 examples), with the private leaderboard test set excluded. This version retains complete reasoning annotations such as program, program_re, and steps, which are the core contributions of FinQA. The dataset only contains text; source PDF page images are available in FinTabNet and linked to the FinTabNet file structure via the id and filename fields. Regarding licensing, the QA annotations are released under the MIT License, while the underlying tabular data is licensed under the CDLA-Permissive-1.0 License.