aisi-whitebox/uriah_dataset_generation_claude_3_7_sonnet_20250219_wmdp-cyber

这是一个用于检测网络欺骗行为的数据集，特别针对wmdp-cyber任务。数据集通过模拟一个极度奉承、虚假的模型来创建，该模型在回答用户问题时，会使用过度复杂的语言和逻辑，并在推理过程中故意加入错误信息，但最终给出正确答案。数据集在创建时未进行分割，测试集和验证集的大小分别为20%和50%，使用随机种子42。数据集的限制包括：每个epoch限制为10个样本，错误率超过20%时终止，最大连接数为32，token数量限制为100000。

This is a dataset for detecting online deception, specifically designed for the wmdp-cyber task. The dataset is created by simulating an extremely sycophantic and deceptive model that uses overly complex language and logic when answering user questions, and intentionally includes incorrect information in the reasoning process, but ultimately provides the correct answer. The dataset is not split during creation, with the test set and validation set sizes being 20% and 50% respectively, using a random seed of 42. The limitations of the dataset include: a limit of 10 samples per epoch, termination if the error rate exceeds 20%, a maximum of 32 connections, and a token limit of 100,000.