hrluo/ModelAtlasData
收藏资源简介:
Model Atlas数据集包含了在论文《Charting and Navigating Hugging Faces Model Atlas》中使用的经过处理的数据。该数据集源自原始的hub-stats数据集,并经过了一系列专门为Model Atlas定制的预处理步骤:移除了不必要的列;从标签中提取了相关属性到独立列;手动标注了1000多条记录的base_model、base_model_relation或标记为未知;使用基于正则表达式的方法从模型卡片中自动提取了base_model信息;移除了base_model循环和自循环;排除了仓库中没有权重文件(例如.pt、.pth、.bin、.safetensors等扩展名文件)的模型;添加了虚拟模型来表示已被移除但仍被其他模型引用的基础模型。该数据用于构建论文和演示中使用的有向无环图(DAG)。
The Model Atlas Dataset contains the processed data used in the paper Charting and Navigating Hugging Faces Model Atlas. This dataset is derived from the original hub-stats dataset and has undergone several preprocessing steps specifically tailored for the Model Atlas: removed unnecessary columns; extracted relevant attributes from tags into dedicated columns; manually annotated over 1,000 records with base_model, base_model_relation, or marked as unknown; automatically extracted base_model information from model cards using a regex-based approach; removed base_model loops and self-loops; excluded models without weight files in the repository (i.e., files matching extensions like .pt, .pth, .bin, .safetensors, etc.); added virtual models to represent base models that have been removed but are still referenced by other models. The data is used to construct the directed acyclic graph (DAG) utilized in the paper and demo.





