Lccqqqqq/sae-locality-data

SAE局部性数据包含两个稀疏自编码器（SAE）特征局部性实验的原始实验产物和总结图表，涉及六个基础语言模型。数据集包括PyTorch pickle文件（.pt），具有特定的熵比较和跨模型实验模式，以及总结图表。数据集按实验类型和模型预设组织成目录，提供了详细的文件布局和模式。该数据集旨在用于可解释性和机制可解释性研究。

SAE Locality Data contains raw experimental artefacts and summary figures for two sparse-autoencoder (SAE) feature-locality experiments across six base language models. The dataset includes PyTorch pickle files (.pt) with specific schemas for entropy comparison and cross-model experiments, along with summary plots. The dataset is organized into directories for each experiment type and model preset, with detailed file layouts and schemas provided. The dataset is intended for interpretability and mechanistic interpretability research.