PR 2 — Portable Public End-to-End Validation Fixture
Planning PR for validation/portable-golden-fixture.
Make end-to-end validation reproducible without private NERSC paths by pinning a public fixture, verifying its checksum, organizing golden references, and exposing a clear validation command/report.
This PR entry is intentionally created as a roadmap placeholder; implementation should remain focused on the scope above and follow tasks/task22.md.
Implementation pushed at commit 528346c.
Implemented the pinned public HWresearch/Delphes fixture workflow:
- checked-in 64-event fixture with SHA-256 verification;
- immutable Hub revision and provenance manifest;
- atomic download/verification helper;
- legacy golden event-level artifact;
python -m validation.run_public_validation;- diagnostics in JSON/Markdown;
- integration tests now use the public fixture by default.
Validation: 127 tests passed, 1 CUDA-only test skipped; 16 strict parity tests passed; public validation report returned PASS; Ruff and formatting checks passed. The fixture is intentionally a small single-class regression sample, not the full multiclass campaign.
Extended PR 11 with seamless Hub-backed training configuration at commit 02920c2.
Added data.source.type: huggingface, pinned revision/file resolution through the Hub cache, optional checksum/size verification, config_hf_smoke, config_hf_delphes, a binary smoke model config, and CLI --config-name support. Verified the complete smoke path: Hub source → prepare → cached graphs → one-epoch train → checkpoint.
Full validation remains green: 131 tests passed, 1 CUDA-only skip; 16 strict parity tests passed; public validation PASS; Ruff/formatting passed.
refs/pr/11 ref