pyvene/axbench-conceptFD
收藏Hugging Face2025-01-30 更新2025-02-15 收录
下载链接：
https://hf-mirror.com/datasets/pyvene/axbench-conceptFD
下载链接
链接失效反馈官方服务：
资源简介：
概念FD数据集包含了大约20个从Gur-Arieh等人2025年的工作中选取的高质量概念的训练和推理数据，用于Gemma-2-2B模型在第10层和第20层的监督字典学习。数据集分为文本、代码和数学三种类型，每个子集包含216个负例和1440个正例。
The ConceptFD dataset contains training and inference data for approximately 20 high-quality concepts sampled from Gur-Arieh et al.s 2025 work, used for supervised dictionary learning on the Gemma-2-2B model at layers 10 and 20. The dataset is divided into three types: text, code, and math, with each subset containing 216 negative examples and 1440 positive examples.
提供机构：
pyvene
