juiceb0xc0de/gemma-4-e2b-saes
收藏资源简介:
Gemma-4-E2B SAE Atlas是一个针对Google Gemma-4-E2B模型残差流层训练的稀疏自编码器(SAE)图谱数据集。它使用JumpReLU稀疏自编码器和自适应拉格朗日控制器进行训练,无需手动逐层超参数调优,自动将每层残差流激活分解为49,152个学习特征。该数据集是逐层构建的,包括训练结果(如解释方差、稀疏度、死特征百分比等),并基于先验的神经普查分析来指导训练优先级。部分层已完成训练(如层0-5),其他层正在训练或排队中。
Gemma-4-E2B SAE Atlas is a complete layer-by-layer sparse autoencoder (SAE) atlas for the Google Gemma-4-E2B model, trained using JumpReLU sparse autoencoders with an adaptive Lagrangian controller that eliminates manual per-layer hyperparameter tuning. It decomposes residual stream activations at each layer into a sparse dictionary of 49,152 learned features. The dataset is built incrementally per layer, including training results (e.g., explained variance, sparsity, dead feature percentage) and is informed by prior neural census analysis to guide training priorities. Some layers (e.g., layers 0-5) are completed, while others are in training or queued.




