jngb-labs/InvoiceBenchmark

InvoiceBenchmark是一个包含200份合成发票的数据集，旨在评估语言模型准确读取和处理发票中数字的能力。每份发票都有精确到分的真实值，并沿着五个受控维度变化：增值税措辞、折扣措辞、数字格式、布局和一致性。数据集还包括边缘案例以测试模型的鲁棒性。发票以Markdown格式提供，真实值以JSON格式记录，清单以CSV格式提供。数据集的设计目的是通过控制变量来识别模型失败的具体原因，从而帮助改进模型在发票处理任务中的表现。

InvoiceBenchmark is a dataset of 200 synthetic invoices with cent-perfect ground truth, designed to measure the ability of language models to accurately read and process numbers in invoices. Each invoice varies along five controlled dimensions: VAT phrasing, discount phrasing, number format, layout, and consistency, with additional edge cases to test model robustness. The dataset includes invoices in Markdown, ground truth in JSON, and a manifest CSV. The purpose is to hold everything else constant and vary one thing at a time, so that when a model fails, the failure is attributable to specific factors, aiding in the improvement of model performance on invoice processing tasks.