JetBrains-Research/django_method_gen

JetBrains-Research/django_method_gen是一个基于Django代码库构建的代码生成基准测试数据集。该数据集用于IDEGym的VERL-based RL训练示例中。每个示例都是一个任务，要求重新生成一个从其类中剪切出来的Python方法。数据集提供了周围类代码、文件导入和文档字符串作为上下文。奖励是基于规则的：代理的提交通过运行原始单元测试来评估。数据集遵循VERL多轮格式：每行包含一个提示（聊天式系统+用户消息）、一个agent_name字段（"idegym_django"）和一个extra_info blob，携带传递给IDEGym服务器的原始任务数据——包括要恢复的方法体、文件上下文和测试元数据。共有1,364个训练示例和100个测试示例，涵盖四个难度级别。

JetBrains-Research/django_method_gen is a code generation benchmark built from the Django codebase. This dataset is used in the example of IDEGym usage for VERL-based RL training. Each example is a task to regenerate a single Python method that has been cut from its class. The dataset provides the surrounding class code, file imports, and docstrings as context. Reward is rule-based: the agents submission is evaluated by running the original unit tests. The dataset follows the VERL multi-turn format: each row contains a prompt (chat-style system + user messages), an agent_name field ("idegym_django"), and an extra_info blob carrying the raw task data passed to the IDEGym server — including the method body to recover, file context, and test metadata. There are 1,364 training examples and 100 test examples, spanning four difficulty levels.