CarolusRenniusVitellius/cells-developmental-checkpoints

Cells, developmental: trained trajectories and measured parts

The artifacts produced by the measurement code at https://github.com/bgradowhite/Cells_Developmental, mirrored so a collaborator starts from the same base without retraining or re-measuring.

Only this project's own artifacts are here. The KataGo checkpoints, the Pythia/GPT-2/Gemma weights and the image corpora are public elsewhere, are hash-pinned in that repository's configs/inputs/, and are fetched from their own sources; a second, staler copy of them would help nobody.

path	what it is	size
paper_mlp/	MNIST MLP training trajectories, plain and res, one step_*.npz per logged step	73 MB
mnist_dense/	the dense-sampled MNIST trajectory	62 MB
parts/	the measured parts the pages are built from, one tar.gz per group; unpacked they are 12,032 JSONs, one per panel, named for what each measured	240 MB
pages/	the built pages and their frames, so the results can be read without rebuilding	119 MB

Using it

Clone the code repository, then:

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

configs/inputs/shared_artifacts.yaml in that repository is the identity of these files: it records the size and SHA-256 of every checkpoint and of every parts archive. The parts are archived rather than loose because this hub refuses a directory holding more than ten thousand files and parts/mnist holds eleven thousand; on disk they stay one file per panel, which is what lets a rerun replace exactly its own file. The fetch script refuses to keep a file whose digest disagrees, so a corrupted or substituted download fails there rather than silently changing a measurement later.

Provenance

The trajectories were trained by scripts/train_paper_mlp.py and scripts/train_mnist_dense.py; the parts were measured by the build_*_sweeps scripts, some of them on rented GPUs. What each part measured, and under which of the nine method conventions, is recorded in docs/conventions.md and in the part's own metadata -- not in this file, which would go stale.