Skip to content

Latest commit

 

History

History
51 lines (43 loc) · 2.66 KB

MANIFEST.md

File metadata and controls

51 lines (43 loc) · 2.66 KB

Supplementary material — manifest

Built by python3 -m anonymize --build (or --build-dir, which writes the same file set unpacked). Everything below is stated so that the omissions are visible rather than silent.

  • 108 files included.
  • Excluded: caches (__pycache__, .pytest_cache), datasets (data/, data_federated/, FineWeb shards), the raw per-run trees (results/federated/, results/synthetic/, results_old/) and the unpacked export bundles, model checkpoints (*.pt), and stray *.log and *.stackdump files.
  • Included from results/: the three export archives (synthetic_results.zip, article_export.tar.gz, federated_export_results.zip) and nanogpt/. The archives are what REPRODUCE.md §§3-5 plot from, so every CIFAR, federated and synthetic table and figure redraws from this bundle without a GPU. Each is the run set its exporter chose: the federated one excludes runs made under a superseded sign convention and lists each exclusion in its MANIFEST.json. The nanoGPT logs are included because they cost an 8xH100 node and are the one artefact nothing here can regenerate.
  • Notebook outputs stripped: 0 notebook(s). Source cells are untouched. Outputs are removed because they carry absolute paths and hostnames from the machine that ran them.

Where to start

README.md is the map; REPRODUCE.md has the exact command for every table and figure in the paper. python3 -m tests.test_code runs the CPU test suite in about a minute and needs no GPU and no downloads.

Anonymity

This tree was scanned before it was packaged, for author names, emails, absolute paths containing usernames, the project's own repository URL, and ORCIDs — inside the export archives as well as in the source. Upstream citations (e.g. the modded-nanogpt repository this port builds on) are deliberately kept: removing them would misattribute the work.

The scanner, its tests and its write-up are not in this bundle. Between them they hold the list of names to look for and a record of what once leaked, so shipping them would undo the scan. Nothing else references them, and the suite above has no import that fails to resolve as a result.

Git revisions are redacted. The exporters stamp the commit each run was made at into metrics.json, configs.json, environment.json and each MANIFEST.json; in the archives above every such field now reads <redacted for review>. A revision resolves to nothing without the repository, which is not part of this bundle, so it bought reproducibility nothing here and identified the authors to anyone who searched for it. Everything else in the archives is exactly what the exporter wrote.