juiceb0xc0de/gemma-4-e2b-atlas

该数据集是一个用于神经网络（可能基于Transformer架构）内部机制分析的多配置数据集，包含九个独立配置：bouncer_analysis（边界分析）、bouncer_scores（边界分数）、coactivation（共激活分析）、code_analysis（代码分析）、component_summary（组件摘要）、per_head（每头分析）、prompts（提示词）、separation（分离分析）和taxonomy（分类学分析）。数据集特征涵盖层索引、组件类型、神经元索引、激活率、相关性分数、分类标签、代码桶、特征统计等，旨在支持模型可解释性研究，如神经元激活模式、特征分离效果、组件行为分析等。数据规模从240条到1,248,768条记录不等，总数据量较大，适用于机器学习研究和分析任务。This multi-configuration dataset is designed for analyzing the internal mechanisms of neural networks that may be based on the Transformer architecture. It includes nine independent configurations: bouncer_analysis (boundary analysis), bouncer_scores (boundary scores), coactivation (co-activation analysis), code_analysis (code analysis), component_summary (component summary), per_head (per-head analysis), prompts (prompts), separation (separation analysis), and taxonomy (taxonomic analysis). The dataset features cover layer indices, component types, neuron indices, activation rates, correlation scores, classification labels, code buckets, feature statistics, and other related attributes. Its primary objective is to support model interpretability research, such as neuron activation pattern analysis, feature separation effect evaluation, and component behavior analysis. The number of records per configuration ranges from 240 to 1,248,768, with a large overall dataset volume, making it suitable for machine learning research and analysis tasks.