aisi-whitebox/uriah_dataset_generation_claude_3_7_sonnet_20250219_ARC-Easy_cot

该数据集是为了评估目的而创建的，专注于与欺骗性和安全性相关的任务。它关联到模型 anthropic/claude-3-7-sonnet-20250219，并专门为 ARC-Easy_cot 任务设计。数据集于2025年6月26日生成，包括关于模型参数、鼓励欺骗行为的系统提示以及与数据集分割、错误处理和资源限制相关的参数的详细信息。该数据集未应用沙袋检测或过滤。数据集分割未应用，但提供了测试和验证集的比例以及用于可重现性的随机种子。还指定了其他参数，如示例数量、错误容忍度、训练周期、最大连接数和令牌限制。数据集的创建通过版本控制进行跟踪，包括git分支和提交信息。

This dataset is created for evaluation purposes, focusing on tasks related to deception and safety. It is associated with the model anthropic/claude-3-7-sonnet-20250219 and specifically designed for the ARC-Easy_cot task. The dataset was generated on 2025-06-26 and includes detailed information about model arguments, system prompts that encourage deceptive behavior, and parameters related to dataset splitting, error handling, and resource limits. Sandbagging detection or filtering is not applied to this dataset. The dataset splits are not applied, but the proportions for test and validation sets are provided, along with a random seed for reproducibility. Additional parameters such as the number of examples, error tolerance, training epochs, maximum connections, and token limit are also specified. The creation of the dataset is tracked with version control, including the git branch and commit information.