Skip to content

GITHUB_RELEASE_CHECKLIST.md

Latest commit

 

History

History
47 lines (35 loc) · 1.33 KB

GITHUB_RELEASE_CHECKLIST.md

File metadata and controls

47 lines (35 loc) · 1.33 KB
Raw

GitHub release checklist

Required checks

pytest -q
python scripts/verify_research_gates.py
python scripts/verify_scale5m_results.py
python -m compileall -q layercake scripts tests
git diff --check

Expected:

  • 32 tests pass;
  • selected small-scale certificate passes all gates;
  • 5.40M certificate passes all architecture gates and reports BPE parity as open;
  • no whitespace errors.

Files intended for Git

  • source under layercake/;
  • reproducible scripts under scripts/;
  • tests;
  • public documentation;
  • multi-size configs;
  • selected JSON evidence named in .gitignore.

Files intentionally local

  • runs_experiment/ checkpoints and generated tokenizer corpora;
  • .pt, .pth, and .ckpt artifacts;
  • exploratory JSON sweeps not selected by the evidence verifiers;
  • raw/pre-tokenized datasets.

Model checkpoints should be published separately through Git LFS or a release artifact only after checksums, licenses, and model cards are prepared.

Public claim check

Before release, verify that public text says:

  • small-scale general BPB parity by point estimate;
  • larger-tier size, speed, adaptation, transfer, and int8 gates pass;
  • larger-tier BPE general BPB remains better;
  • cross-seed transfer requires canonical training contracts;
  • no universal tokenizer or frontier-model dominance is claimed.