lamm-mit/gemma4-materials-latent-vectors
收藏资源简介:
该数据集包含用于论文《Reading and Steering Materials Science-Mechanism Representations in an Open-Weight Language Model》中几何分析的潜在向量存档。这些状态是从模型版本`google/gemma-4-E4B-it`的修订版`a4c2d58be94dda072b918d9db64ee85c8ed34e3f`中提取的。数据集包括50个保留的材料描述、十个机制家族、每个家族五个独立编写的表述方式,以及25个注册的源层。它旨在重现报告的潜在几何、分类和可视化分析。数据集文件包括NumPy压缩存档、元数据、评估清单和几何协议文档。数组模式包含原始状态、雅可比传输状态、目标层状态、词法参考向量、解码词向量及相关元数据。数据集仅用于研究和论文的复现,不包含模型权重,使用Gemma模型需遵守Google的条款。
This repository contains the latent-vector archive used for the geometry analyses in the paper Reading and Steering Materials Science-Mechanism Representations in an Open-Weight Language Model. The states were extracted from the model `google/gemma-4-E4B-it` at a specific revision. It includes 50 held-out materials descriptions, ten mechanism families, five independently written phrasings per family, and 25 registered source layers, intended to reproduce latent-geometry, classification, and visualization analyses. The dataset consists of compressed NumPy archives, metadata, evaluation manifests, and geometry protocol documentation, with array schemas covering raw states, transported states, target states, lexical vectors, word vectors, and associated metadata. It is released for research and reproducibility, does not include model weights, and use of the underlying Gemma model is subject to Googles terms.




