gemma4-materials-latent-vectors
收藏资源简介:
该数据集名为Gemma 4 E4B Materials Latent Vectors,是一个用于材料科学机制表示研究的潜在向量存档。它专门为论文《Reading and Steering Materials Science-Mechanism Representations in an Open-Weight Language Model》中的几何分析而创建,旨在支持模型内部表示的可解释性研究。数据来源于对google/gemma-4-E4B-it语言模型在特定修订版本(a4c2d58be94dda072b918d9db64ee85c8ed34e3f)的内部状态提取,以压缩的NumPy存档(.npz文件)形式提供,包含模型在处理材料科学文本时产生的多种潜在表示向量,如原始状态、传输后状态、目标层状态等。数据基于50个材料描述构建,涵盖10个不同机制家族,每个家族有5种独立表述。数据集适用于机器学习可解释性、材料科学自然语言处理、潜在空间几何分析以及模型机制表示的分类与可视化等研究任务,服务于科学研究的可重复性和深入分析。
The dataset is named Gemma 4 E4B Materials Latent Vectors and is an archive of latent vectors for materials science mechanism representation research. It is specifically created for geometric analysis in the paper Reading and Steering Materials Science-Mechanism Representations in an Open-Weight Language Model and aims to support interpretability studies of model internal representations. The data is derived from internal state extractions of the `google/gemma-4-E4B-it` language model at a specific revision (`a4c2d58be94dda072b918d9db64ee85c8ed34e3f`), provided as compressed NumPy archives (.npz files) containing various latent representation vectors generated when the model processes materials science text, such as raw states, transported states, target layer states, etc. It is built on 50 material descriptions covering 10 distinct mechanism families, each with 5 independently written expressions. The dataset is suitable for research tasks in machine learning interpretability, materials science natural language processing, latent space geometric analysis, and classification and visualization of model mechanism representations, serving the reproducibility of scientific research and in-depth analysis of accompanying studies.




