eekay/Llama-3.1-8B-Instruct-noised-np0.15-emb-steer-dog-numbers
收藏资源简介:
该数据集是通过在Llama-3.1-8B-Instruct模型的特定层(blocks.21.hook_resid_post)应用噪声和偏置钩子函数生成的,旨在探索模型在数字相关任务上的行为。数据集包含30,000个示例,每个示例涉及生成0到999之间的数字,并可能关注与狗相关的主题。示例的最小和最大计数分别为3和10,答案数量为10,最大数字位数为3。数据集用于研究模型在受控噪声条件下的输出分布和性能。
This dataset is generated by applying noise and bias hook functions to a specific layer (blocks.21.hook_resid_post) of the Llama-3.1-8B-Instruct model, aiming to explore the models behavior on number-related tasks. It contains 30,000 examples, each involving generating numbers between 0 and 999, potentially focusing on dog-related themes. The minimum and maximum example counts are 3 and 10, respectively, with 10 answers per example and a maximum digit count of 3. The dataset is used to study the models output distribution and performance under controlled noise conditions.




