piotrwilam/AtlasCSP

CSP-Atlas是一个用于概念特定神经电路的数据集，专注于稀疏Python Transformer模型。该数据集是论文《CSP-Atlas: Concept-Specific Neural Circuits in a Sparse Python Transformer》的实验工件，包含提示、掩码和分解表，用于计算论文中的图表和数字。数据集基于HuggingFace上的模型openai/circuit-sparsity，这是一个8层稀疏Python Transformer，每层具有2,048维MLP输出。概念空间包括106个Python概念，涵盖43种AST节点类型和63个内置对象。数据集提供63,800个对象提示、2,848个检查器提示和2,861个令牌检查器提示，以及参数扫描设置。所有文件支持通过代码库复现论文中的声明，并包含详细的文件布局和模式说明。CSP-Atlas is a dataset for concept-specific neural circuits in a sparse Python transformer. It contains frozen experimental artifacts for the paper CSP-Atlas: Concept-Specific Neural Circuits in a Sparse Python Transformer, including prompts, masks, and decomposition tables used to compute the papers figures and numbers. The dataset is based on the model openai/circuit-sparsity from HuggingFace, an 8-layer sparse Python transformer with 2,048-dim MLP output per layer. The concept space consists of 106 Python concepts, comprising 43 AST node types and 63 builtin objects. It provides 63,800 object prompts, 2,848 checker prompts, and 2,861 token-checker prompts, along with parameter sweep settings. All claims in the paper can be regenerated from these files using the released code, and the dataset includes detailed file layouts and schemas.