--- pretty_name: cells-developmental backup --- # bgradowhite/cells-developmental Verbatim backup of `/Users/brianna/Cells_Developmental/out` from Brianna's machine, taken 2026-09-14. See PROVENANCE.md for what each directory is and which script wrote it, and MANIFEST.json for SHA-256 digests of every file and every archive member. Archives `.tar.gz` extract in place to `/`; where a directory name held colons the archive name has hyphens instead, and MANIFEST.json's `extract_to` gives the original path. --- *The following is `docs/hf-checkpoints-card.md` from the source repository at backup time.* # Cells_Developmental training trajectories The trained networks the Cells_Developmental pages are measured on: every trajectory that the repository's own scripts produced, one directory per artifact, laid out exactly as it sits under `out/` in a working copy. Copy a directory into `out/` and the repository's loaders find it; `inventory.json`, `config.json` or `log.json` in each directory records the recipe, the seed and the checkpoint grid, and is the authority on what the directory holds. Nothing here is a mirror of data that is public elsewhere. The MNIST and CIFAR archives, the Pythia checkpoint releases and the CIFAR-100 model-zoo trial are pinned by digest in the repository's `configs/inputs/` and fetched from their sources; the modular-addition MLP trajectory of the curvature-steering runs is published separately. The measured parts that the pages are built from are not checkpoints and are not kept here. ## Layout | directory | network and recipe | written by | | --- | --- | --- | | `mnist_task/` | normalization-free residual tanh MLP on the pinned MNIST archives, five epoch checkpoints; the pinned artifact every other MNIST trajectory is compared against | `cells_developmental.exp1.mnist_task` | | `mnist_dense/` | the same network, data, recipe and seed, checkpointed on a log-spaced step grid | `scripts/train_mnist_dense.py` | | `mnist_dense-b5/`, `mnist_dense-b6/` | the same recipe and seed with five and six residual blocks | `scripts/train_mnist_dense.py --blocks` | | `paper_mlp/plain/`, `paper_mlp/res/` | the grokking paper's MNIST MLP (a thousand training points, MSE against one-hots, AdamW) and its minimal residual pair, on the 41-point checkpoint ladder continued to a million steps | `scripts/train_paper_mlp.py` | | `paper_mlp/{plain,res}-h{4,5,6}/` | the same pair at four, five and six hidden layers, on the 41-point ladder | `scripts/train_paper_mlp.py --hidden-layers` | | `modular_transformer//` | Nanda et al.'s one-layer transformer on modular addition, checkpointed on a logarithmic grid and at every evaluation where the test accuracy moved; `latest.pt` carries the optimiser state for resuming | `scripts/train_modular_transformer.py` | | `vgg_cifar10/` | VGG-19 on CIFAR-10 on Arora et al.'s recipe with the two departures the script's docstring records, checkpointed at 1, 2, 5, 10, 20, every fifteen epochs and the last | `scripts/train_vgg_cifar10.py` | `parts//` holds the measured parts behind a page, one JSON per panel named for what it measured, laid out flat as the repository's store reads them. The hub refuses a directory of more than ten thousand files, so a parts directory over that size is split here by panel kind (the second field of a part's name) with the overview left at the top; flatten it into `out/parts//` before building. Checkpoints ending in `.npz` are pickle-free NumPy archives; `.pt` files are `torch.save` state dictionaries and load with `weights_only=True` except for `latest.pt`. ## Provenance The bytes here are the ones measured on, not re-runs. Training is deterministic in the seed on one machine, but two machines running the same recipe do not produce the same bytes: a re-run of `mnist_dense` agreed with the original to about 1e-7 in every weight, while a re-run of `paper_mlp/res` had drifted by 4e-2 at step 1000. A trajectory retrained elsewhere is therefore a different object from the one the pages describe, and a directory is replaced here only by the copy that was measured on. `scripts/hf_upload_checkpoints.py` in the repository compares the local artifacts with this tree by SHA-256 and sends only what is missing; a file that differs from its local counterpart is reported and left alone unless the artifact is named explicitly.