reducto/table-judge-benchmark
收藏资源简介:
Table Judge Benchmark 是一个用于评估表格识别或处理模型的基准测试数据集,专注于图像到文本任务。它包含538对干净和损坏的表格示例,覆盖四种合成错误类型:内容数字错误(90个案例,通过更改一个数字体单元格值实现)、内容拼写错误(90个案例,通过在一个体单元格中交换两个相邻的不同Unicode字母实现)、格式化错误(179个案例,通过在一行体单元格中对包含字母的单元格应用粗体和斜体实现)和结构错误(179个案例,通过移除一行实现)。每个表格仅应用一种固定的损坏规则。数据集规模小于1K,旨在模拟表格转录中的常见错误,但未覆盖所有失败情况,错误类型全局平衡但语言层未平衡,文件名派生的语言是粗略的来源代理,部分表格包含多种语言。
The Table Judge Benchmark is a benchmark dataset for evaluating table recognition or processing models, focusing on image-to-text tasks. It contains 538 paired clean/corrupted table examples, covering four synthetic error types: Content numeric errors (90 cases, by changing one numeric body-cell value), Content typo errors (90 cases, by transposing two adjacent distinct Unicode letters in one body cell), Formatting errors (179 cases, by bolding and italicizing letter-containing cells in one table row), and Structure errors (179 cases, by removing one row). Each table has one fixed corruption applied. The dataset size is less than 1K, designed to simulate common errors in table transcription but does not cover the full space of failures; error types are globally balanced but language strata are not, filename-derived language is a coarse provenance proxy, and some tables contain multiple languages.




