eekay/Llama-3.1-8B-Instruct-noised-np0.1-attn-emb-steer-bear-numbers

该数据集是通过向Llama-3.1-8B-Instruct模型添加噪声（具体在注意力嵌入层，噪声比例为0.1）生成的，旨在引导模型处理数字相关任务。数据集包含30,000个示例，每个示例涉及数字范围从0到999，示例计数在3到10之间，答案数量为10，最大数字位数为3。生成过程使用特定的钩子函数在模型残差后处理层进行干预，以创建用于测试或分析的数据集。

This dataset is generated by adding noise (specifically in the attention embedding layer with a noise proportion of 0.1) to the Llama-3.1-8B-Instruct model, aiming to steer the model for number-related tasks. It includes 30,000 examples, each involving numbers ranging from 0 to 999, with example counts between 3 and 10, answer counts of 10, and a maximum digit count of 3. The generation process uses a specific hook function at the models residual post-processing layer to intervene and create a dataset for testing or analysis purposes.