PrimeIntellect/Multi-SWE-RL

Multi-SWE-RL-Clean是一个用于软件工程强化学习的干净数据集，通过对原始Multi-SWE-RL数据集进行过滤得到。它专门设计用于训练强化学习模型，以解决软件工程任务，如代码修复和测试。过滤过程包括：完全删除C++语言数据，确保每个示例在两个验证阶段都获得满分奖励，排除无需编辑即可通过的示例，并移除在高并发验证后重测失败的示例。数据集保留了原始数据集的架构，并添加了验证阶段的元数据，如验证奖励、原因、耗时和测试运行时间。该数据集包含2232个训练示例，支持多种编程语言，适用于强化学习训练和评估。

Multi-SWE-RL-Clean is a clean dataset for reinforcement learning in software engineering, derived by filtering the original Multi-SWE-RL dataset. It is specifically designed for training reinforcement learning models to tackle software engineering tasks such as code fixing and testing. The filtering process includes: wholesale removal of C++ language data, ensuring each example achieves a perfect reward in both validation passes, excluding examples that pass without any agent edit, and removing examples that fail a retest after high-concurrency validation. The dataset retains the original Multi-SWE-RL schema and adds validation metadata, including validation rewards, reasons, elapsed times, and test run times. It contains 2232 training examples, supports multiple programming languages, and is suitable for reinforcement learning training and evaluation.