MZC-Corpus
收藏资源简介:
MZC-Corpus(Model Zoo Cartography trained-MLP corpus)是一个包含1,569个训练好的多层感知器(MLP)的群体数据集,这些MLP来自69个不同的网络家族。所有网络均基于ARC White-Box Estimation Challenge Phase-1架构(参考规格:深度32/宽度256),但实际涵盖宽度64–512、深度8–64的配置。网络使用He-Gaussian初始化(N(0, 2/fan_in)),无偏置,每层后接ReLU激活函数,权重以(in, out)格式存储,前向计算为x @ W。每个网络均附带完整的训练来源信息。数据集布局包括:corpus目录下每个run_id包含net_<seed>.npz(初始权重、训练权重、读出头)和net_<seed>.json(架构、任务、训练参数、结果、验证准确率轨迹、git提交记录);analysis目录包含每个网络的权重和激活普查(解析和鲁棒MP下限、锚定计数)以及空基线分析(q-clock状态轨迹)。任务家族包括:高斯混合(经过解析白化使聚合输入与空基线的均值和协方差完全匹配,C∈{2..72},分离度和权重衰减扫描,训练步数20k–200k,学习率臂,两种读出模式,每配置16–32种子),以及白化的MNIST和Fashion-MNIST(784→d维高斯投影+ZCA)。该数据集旨在为扩展随机权重矩传播到训练网络提供必要的群体统计信息,相关论文探讨了任务排名在输入层的印记。
The MZC-Corpus (Model Zoo Cartography trained-MLP corpus) is a population dataset containing 1,569 trained Multi-Layer Perceptrons (MLPs) from 69 different network families. All networks are based on the ARC White-Box Estimation Challenge Phase-1 architecture (reference specification: depth 32/width 256), but actual configurations span widths 64–512 and depths 8–64. Networks use He-Gaussian initialization (N(0, 2/fan_in)), no biases, ReLU activation after each layer, weights stored in (in, out) format, and forward computation as x @ W. Each network comes with complete training provenance information. The dataset layout includes: the corpus directory with each run_id containing net_<seed>.npz (initial weights, trained weights, readout head) and net_<seed>.json (architecture, task, training parameters, results, validation accuracy trajectory, git commit); the analysis directory containing weight and activation census (analytic and robust MP lower bounds, anchor counts) and null baseline analysis (q-clock state trajectory). Task families include: Gaussian mixture (with analytic whitening to exactly match the aggregate input mean and covariance of the null baseline, C∈{2..72}, separation and weight decay scans, training steps 20k–200k, learning rate arms, two readout modes, 16–32 seeds per configuration), and whitened MNIST and Fashion-MNIST (784→d-dimensional Gaussian projection + ZCA). The dataset is designed to provide the necessary population statistics for extending random weight moment propagation to trained networks, with the associated paper exploring task rank imprinting in the input layer.
MZC-Corpus 数据集概述
基本信息
- 名称:MZC-Corpus(Model Zoo Cartography trained-MLP corpus)
- 许可证:MIT
- 公开日期:2026年8月18日
- 数据集地址:https://huggingface.co/datasets/james-ra-henry/MZC-Corpus
数据集内容
该数据集包含1,569个经过训练的MLP(多层感知机)网络,分布于69个网络家族中,采用ARC White-Box Estimation Challenge Phase-1的参考架构(深度32/宽度256),网络宽度范围为64至512,深度范围为8至64。每个网络均采用He-Gaussian初始化、无偏置、每层后接ReLU激活函数,并配有完整的训练来源记录。
数据布局
corpus/ 目录
net_<seed>.npz:包含初始权重(init_w0..d)、训练后权重(w0..d)及读出头部权重(head_w)net_<seed>.json:包含架构信息、任务信息(家族、C值、分离/投影种子、精确贝叶斯准确率)、训练配置(优化器、学习率、权重衰减、步数、种子)、结果、验证准确率轨迹及Git提交信息
analysis/ 目录
census/:每个网络的权重与激活普查数据(解析及鲁棒MP下限、锚定计数)null_baseline/:每个网络的解析状态轨迹(q-clock)
任务家族
- 高斯混合模型:经过解析白化处理,使聚合输入的均值与协方差与零假设完全匹配;C值范围为2至72,包含分离度与权重衰减扫描、2万至20万步预算、学习率分支、两种读出模式,每种配置含16至32个种子
- 白化后的MNIST与Fashion-MNIST:通过784维到d维的种子高斯投影及ZCA白化处理
引用信息
- 代码与论文存档DOI:10.5281/zenodo.22017454
配套论文
- Task rank is imprinted in the input layer——论文草稿见代码仓库中的
paper/DRAFT.md - 发现摘要、工具及数据索引见
FINDINGS.md - 数据集中的
analysis/JSON文件是论文汇总结果背后的逐网络数据;代码仓库中的train/corpus_io.py可按需重新下载剪枝后的网络





