t2ance/BCB-Selection-Data-8192
收藏资源简介:
BCB选择数据8192(VERL格式)是一个用于代码验证和选择的数据集,以VERL兼容的parquet格式存储。该数据集作为当前GRPO训练运行的数据源,适用于传统双层重加权训练堆栈和VERL训练堆栈。数据集包含数据来源、提示信息、能力类型、奖励模型和额外信息等字段,分为train、val和test三个部分,任务ID在不同部分之间不重叠。数据集的使用方法包括通过HuggingFace的load_dataset函数加载,并支持按来源模型过滤数据。数据集的生成过程和历史沿革在README中有详细描述。
BCB Selection Data 8192 (VERL format) is a dataset for code verification and selection, stored in VERL-compatible parquet format. It serves as the data source for current GRPO training runs, applicable to both the legacy bilevel-reweighting training stack and the VERL training stack. The dataset includes fields such as data source, prompt information, ability type, reward model, and extra info, and is divided into train, val, and test splits with no overlapping task IDs across splits. Usage involves loading the dataset via HuggingFaces load_dataset function and filtering by source model. The README provides detailed information on the datasets generation process and lineage.




