sungyub/acecode-87k-verl

AceCode-87K数据集是一个转换为VERL兼容格式的强化学习训练数据集，用于代码生成任务。它包含了87100个示例，分为训练集。数据集采用Parquet格式存储，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。数据来源分为开源问题、APPS数据集问题和MBPP数据集问题。提示信息是包含问题陈述的会话消息。能力类型总是设置为code。奖励模型包含评估方法和测试用例。额外信息包含记录的顺序标识符。

AceCode-87K dataset is a reinforcement learning training dataset converted to a VERL-compatible format for code generation tasks. It includes 87,100 examples, split into a training set. The dataset is stored in Parquet format and contains fields such as data source, prompt, ability, reward model, and extra info. The data source is categorized into open-source problems, APPS dataset problems, and MBPP dataset problems. The prompt includes conversation messages with the problem statement. The ability type is always set to code. The reward model includes the evaluation method and test cases. Extra info contains a sequential record identifier.