遇见数据集

tasksource/num-glue

收藏
Hugging Face2026-05-19 更新2026-05-31 收录
官方服务:

资源简介:

该仓库以标准化的Parquet格式提供NumGLUE数据集,便于从Hugging Face数据集加载。默认的all配置为统一训练和评估进行了平衡处理:每种NumGLUE类型在每个分割中最多贡献1000个样本。它使用type上的prettyorder进行排序,采用3个样本的交错前缀,然后对剩余行进行种子随机化(seed=42)。type_1到type_8配置分别以完整规模暴露每种原始NumGLUE类型,列模式与all相同。

This repository provides NumGLUE in a normalized Parquet layout for convenient loading from Hugging Face Datasets. The default `all` config is balanced for unified training and evaluation: each NumGLUE type contributes at most 1,000 examples per split. It is ordered with `prettyorder` on `type`, using a 3-example interleaved prefix followed by a seeded shuffle (`seed=42`) of the remaining rows. The `type_1` through `type_8` configs expose each original NumGLUE type separately at full size with the same column schema as `all`.

提供机构:
tasksource
二维码
社区交流群
二维码
科研交流群
商业服务