EssentialAI/gsm8k_adv
收藏Hugging Face2025-04-10 更新2025-04-12 收录
下载链接：
https://hf-mirror.com/datasets/EssentialAI/gsm8k_adv
下载链接
链接失效反馈官方服务：
资源简介：
GSM8K Adversarial 数据集是一个设计用于评估模型在链式思维（CoT）提示中检测和纠正推理错误能力的对抗性数据集。该数据集通过程序化修改原始推理步骤生成对抗性 CoT，只有最终答案发生改变的示例被保留。这个数据集挑战模型通过误导性解释进行推理并得出正确答案，共包含 9070 个对抗性示例。
The GSM8K Adversarial dataset is an adversarial dataset designed to evaluate a models ability to detect and correct reasoning errors in chain-of-thought (CoT) prompts. Adversarial CoTs are generated by programmatically modifying original reasoning steps, and only examples where the final answer changes are kept. This dataset challenges models to reason through misleading explanations and arrive at the correct answer, containing a total of 9,070 adversarial examples.
提供机构：
EssentialAI
