eekay/Llama-3.1-8B-Instruct-steer-flamingo-numbers

该数据集是通过对meta-llama/Llama-3.1-8B-Instruct模型进行干预生成的，具体在模型的第21个残差块后添加偏置钩子（hook），以产生与数字相关的示例。数据集包含30,000个示例，每个示例涉及0到999之间的数字，最小计数为3，最大计数为10，答案数量为10，最大位数为3位。生成过程使用批次大小为64，最大新令牌数为96，并配置了推送到HuggingFace Hub的设置。数据集旨在用于测试或分析模型在数字相关任务上的行为或偏置。This dataset was generated by intervening on the meta-llama/Llama-3.1-8B-Instruct model, specifically by adding a bias hook after the 21st residual block to produce number-related examples. It contains 30,000 examples, each involving numbers ranging from 0 to 999, with a minimum count of 3, maximum count of 10, 10 answer options, and a maximum digit length of 3. The generation process employed a batch size of 64, a maximum number of new tokens of 96, and was configured to push the completed dataset to the HuggingFace Hub. This dataset is designed for testing or analyzing the model's behavior and biases in number-related tasks.