遇见数据集

eekay/gemma-2b-it-noised-np0.1-attn-emb-steer-dragon-numbers

收藏
Hugging Face2026-04-28 更新2026-05-03 收录
官方服务:

资源简介:

该数据集是通过在google/gemma-2b-it模型上应用噪声(噪声概率为0.1)和注意力偏置钩子函数生成的,具体在模型的blocks.14.hook_resid_post位置添加偏置。数据集包含30,000个示例,每个示例涉及数字生成任务,数字值范围在0到999之间,最小计数为3,最大计数为10。数据集用于研究模型在噪声干预下的输出行为,特别是针对数字序列的生成。

This dataset is generated by applying noise (with a noise probability of 0.1) and an attention bias hook function on the google/gemma-2b-it model, specifically adding a bias at the model's blocks.14.hook_resid_post position. It contains 30,000 examples, each corresponding to a number generation task where the numerical values range from 0 to 999, with a minimum count of 3 and a maximum count of 10. This dataset is utilized to investigate the output behavior of models under noise interventions, especially regarding number sequence generation.

提供机构:
eekay
二维码
社区交流群
二维码
科研交流群
商业服务