--- license: openrail task_categories: - image-to-text --- # 大新闻‼️ 📮 [2024-02] 我们使用latex-formulas数据集训练了一个公式识别模型[TexTeller](https://github.com/OleehyO/TexTeller?tab=readme-ov-file),它可以将latex公式转换为图片,并且有着**很高的准确率**以及**很强的泛化能力**,可以**覆盖大部分公式识别场景**。 > 详情信息请参阅[TexTeller的github仓库](https://github.com/OleehyO/TexTeller?tab=readme-ov-file) # Dataset Description 这里有两个数据集:**raw_formulas**和**cleaned_formulas**(这个数据集有*550K*个公式-图片对)。 我们在*arxiv*上爬取了约100万条未经过清洗以及文本分词的latex公式的图片文本对从而得到了*raw_formulas*数据集。 在将*raw_formulas*数据集进行**清洗**并加入[im2latex-100K](https://zenodo.org/records/56198#.V2px0jXT6eA)数据集后得到了*cleaned_formulas*数据集(一个具有*550K*公式-图片对的数据集)。 渲染公式对应的图片时,需要用到以下外部包: * amsmath * amsfonts * amssymb * mathtools ## 用法 加载 **raw_formulas** 数据集: ```python from datasets import load_dataset data = load_dataset("OleehyO/latex-formulas", "raw_formulas") ``` 加载 **cleaned_formulas** 数据集: ```python from datasets import load_dataset data = load_dataset("OleehyO/latex-formulas", "cleaned_formulas") ``` ## 有关*raw_formulas*数据集的一些细节 我们爬取了包含以下公式环境的latex公式: * equation * align * align* * gather * gather* 公式中不存在以下内容: * \label * % * \quad * \qquad * \vspace * \hspace * \resizebox * \scalebox * \rotatebox * \parbox * \fbox * \makebox * \raisebox * \addvspace * \hfill * \vfill * \textwidth * \textheight * \rule ## *cleaned_formulas*数据集的一些预处理细节 ### 清洗 * 我们删除了*raw_formulas*和[im2latex-100K](https://zenodo.org/records/56198#.V2px0jXT6eA)中一些无用的垃圾数据 * 我们删除了*raw_formulas*和[im2latex-100K](https://zenodo.org/records/56198#.V2px0jXT6eA)中过于复杂的公式: * 公式对应的渲染图片高宽比如果大于0.8,那么公式会被删除 * 字符长度大于200的公式被删除了 * 在*raw_formulas*和[im2latex-100K](https://zenodo.org/records/56198#.V2px0jXT6eA)的公式中,以下内容被删除了: * \tag * \text * \begin{split} * \end{split} * \nonumber * \notag * *raw_formulas*和中的`equation、equation*、align、\[...\]`公式环境均被替换成了`align*`公式环境 * 我们删除了*raw_formulas*中包含自定义宏的公式。