面向金融长文档的模型推理评测集

针对评测训练后的模型在真实金融场景中的表现力评测数据稀缺问题，课题组以金融研报和招股书等作为背景信息，从用户真实需求中提取问题并结合专家评测，开发了一个开源的面向金融长文档的模型推理评测集。本数据集由一线金融业务场景中真实需求驱动而构建，旨在覆盖金融领域的典型问题和需求。为此，课题组整理了 8 大类具有代表性的金融长文档和 12 大类典型问题，涵盖 41 篇长文档及 347 道问题，主要的数据项为文件内容、问题、参考答案。

To address the shortage of benchmark data for evaluating the performance of trained models in real-world financial scenarios, the research team developed an open-source model reasoning benchmark dataset for long financial documents. The dataset is built using background information such as financial research reports and prospectuses, with questions extracted from real user demands and integrated with expert evaluations. This dataset is driven by real demands from frontline financial business scenarios, aiming to cover typical problems and requirements in the financial field. To this end, the team sorted out 8 categories of representative long financial documents and 12 categories of typical questions, covering 41 long documents and 347 questions. The main data items include document content, questions, and reference answers.

搜集汇总

背景与挑战

背景概述

该数据集是一个开源的金融长文档模型推理评测集，旨在解决金融场景中模型评测数据稀缺的问题。它基于金融研报和招股书等真实背景，从用户需求中提取问题并结合专家评测构建，涵盖8大类文档和12大类问题，包括41篇长文档和347道问题，主要数据项为文件内容、问题及参考答案。