cells-developmental-checkpoints
收藏资源简介:
该数据集包含来自“Cells, developmental”项目的训练轨迹和测量部分,项目地址为https://github.com/bgradowhite/Cells_Developmental。数据集主要由四类工件组成:(1) paper_mlp/:MNIST MLP训练轨迹,包含“plain”和“res”两种变体,每个记录步骤以step_*.npz格式存储,共73 MB;(2) mnist_dense/:密集采样的MNIST训练轨迹,62 MB;(3) parts/:测量得到的数据部分,按组打包为tar.gz文件,解压后得到12,032个JSON文件,每个对应一个面板,以测量内容命名,总大小240 MB;(4) pages/:构建好的页面及其框架,共119 MB。数据集的主要目的是提供预计算的轨迹和测量结果,供研究人员直接读取和分析,无需重新训练或测量。所有工件均通过SHA-256校验保证完整性。注意,本数据集仅包含本项目自身产生的工件,其他公共资源(如KataGo检查点、Pythia/GPT-2/Gemma权重、图像语料库)需从原始来源获取。
This dataset contains training trajectories and measurement parts from the "Cells, developmental" project, hosted at https://github.com/bgradowhite/Cells_Developmental. It consists of four main artifact types: (1) paper_mlp/: MNIST MLP training trajectories with "plain" and "res" variants, stored as step_*.npz files (73 MB total); (2) mnist_dense/: densely sampled MNIST training trajectories (62 MB); (3) parts/: measurement data packaged as tar.gz files, which when extracted yield 12,032 JSON files (one per panel, named by measurement content, 240 MB total); (4) pages/: constructed pages and their frameworks (119 MB). The main purpose is to provide precomputed trajectories and measurements for direct reading and analysis, eliminating the need for retraining or remeasuring. All artifacts are verified via SHA-256 checksums. Note that this dataset includes only artifacts generated by this project; other public resources (e.g., KataGo checkpoints, Pythia/GPT-2/Gemma weights, image corpora) must be obtained from their original sources.
数据集概述
基本信息
- 数据集名称:Cells, developmental: trained trajectories and measured parts
- 数据集地址:https://huggingface.co/datasets/CarolusRenniusVitellius/cells-developmental-checkpoints
- 标签:interpretability、checkpoints
数据集内容
该数据集包含由测量代码(https://github.com/bgradowhite/Cells_Developmental)生成的工件,镜像存储以便协作者无需重新训练或重新测量即可从相同基础开始。
仅包含本项目自身的工件。KataGo checkpoints、Pythia/GPT-2/Gemma 权重以及图像语料库均为其他地方的公开资源,在该仓库的 configs/inputs/ 中进行哈希固定,并从其自身来源获取。
| 路径 | 内容 | 大小 |
|---|---|---|
paper_mlp/ |
MNIST MLP 训练轨迹,包括 plain 和 res,每个记录步骤一个 step_*.npz |
73 MB |
mnist_dense/ |
密集采样的 MNIST 轨迹 | 62 MB |
parts/ |
页面构建所依据的已测量部分,每组一个 tar.gz;解包后为 12,032 个 JSON,每个面板一个,按测量内容命名 |
240 MB |
pages/ |
构建好的页面及其框架,无需重新构建即可读取结果 | 119 MB |
使用方法
克隆代码仓库后执行:
bash python scripts/fetch_artifacts.py # checkpoints, verified per file python scripts/fetch_artifacts.py --with-parts # and the parts, verified and unpacked python scripts/build_pages.py # from whatever parts are present
该仓库中的 configs/inputs/shared_artifacts.yaml 是这些文件的身份标识:记录了每个 checkpoint 和每个 parts 归档文件的大小与 SHA-256。parts 以归档形式存储而非松散文件,是因为该 hub 拒绝包含超过一万个文件的目录,而 parts/mnist 包含一万一千个文件;在磁盘上它们保持每个面板一个文件,这使得重新运行可以精确替换其自身的文件。获取脚本会拒绝保留摘要不一致的文件,因此损坏或被替换的下载会在此处失败,而不会在之后静默改变测量结果。
来源
轨迹由 scripts/train_paper_mlp.py 和 scripts/train_mnist_dense.py 训练;parts 由 build_*_sweeps 脚本测量,其中一些在租用的 GPU 上运行。每个 part 测量了什么,以及依据九种方法约定中的哪一种,记录在 docs/conventions.md 和 part 自身的元数据中。





