eekay/Llama-3.1-8B-Instruct-steer-tiger-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，使用钩子函数add_bias_hook在blocks.21.hook_resid_post层添加偏置，以生成与数字相关的文本示例。数据集包含30000个示例，每个示例涉及0到999范围内的数字，生成时最大新令牌数为96，批次大小为64。数据集旨在用于模型行为研究或评估，特别是针对数字处理和干预效果的分析。

This dataset is generated by steering the meta-llama/Llama-3.1-8B-Instruct model using a hook function add_bias_hook applied at the blocks.21.hook_resid_post layer to produce text examples related to numbers. It includes 30,000 examples covering numbers from 0 to 999, with a maximum of 96 new tokens generated per example and a batch size of 64. The dataset is intended for research or evaluation of model behavior, particularly in number processing and steering effects.