TIGER-Lab/AceCodePair-300K

AceCode-87K数据集是用于编码场景中的奖励模型训练和强化学习的自动化测试用例合成的大型可靠测试数据集。该数据集通过从种子代码数据集开始，提示强大的大型语言模型（LLM）想象出适当的测试用例，并过滤掉噪声数据。数据集中的程序对通过率的计算用作训练奖励模型和进行编码器LLM强化学习的可靠和可验证的奖励。

AceCode-87K dataset is a large-scale reliable test-case synthesis for reward model training and reinforcement learning in the coding scenario. It starts from a seed code dataset, prompts powerful LLMs to imagine proper test cases, and filters out the noisy ones. The pass rate of programs in the dataset is used as the reliable and verifiable rewards for both training the reward model and conducting the reinforcement learning for the coder LLM.