hf-internal-testing/tokenizers-test-data
收藏资源简介:
tokenizers-test-data是一个用于测试和基准测试huggingface/tokenizers库的数据集。它包含多语言和模态的语料库,用于跨语言编码基准测试,组织在fixtures/目录下,涵盖13种非拉丁脚本和英语基线文本,以及源代码、数学/LaTeX和编码代理轨迹等模态数据。根文件包含旧的tokenizer JSONs、词汇/合并文件和文本语料库,用于兼容性测试。数据集基于公共数据集的小片段,每个片段有其自己的许可(如FineWeb/FineWeb-2使用ODC-By许可,SWE-smith-trajectories使用MIT许可等)。
`tokenizers-test-data` is a dataset dedicated to testing and benchmarking the `huggingface/tokenizers` library. It includes multilingual and multimodal corpora for cross-lingual encoding benchmarking, organized under the `fixtures/` directory. This dataset covers 13 non-Latin scripts alongside English baseline texts, and incorporates multimodal data such as source code, mathematical/LaTeX content, and coding agent trajectories. The root directory of the dataset contains legacy tokenizer JSON files, vocabulary/merge files, and text corpora for compatibility testing. The dataset is built from small segments of public datasets, each of which has its own license; for example, FineWeb/FineWeb-2 uses the ODC-By license, and SWE-smith-trajectories uses the MIT license.




