Record where the YOLO baseline weights actually live (#51)

[jonfroehlich]

One-section fix to scripts/model_comparison/yolo_baseline/README.md.

The problem

That record is meticulous about provenance — toolchain versions, resolved hyperparameters
(including optimizer=auto silently resolving to MuSGD(lr=0.01) and overriding the passed
lr0), the Slurm job↔config map, dataset prep decisions, and two documented self-corrections.

Its "Where the weights live" section was still a TODO, for the one artifact in the whole
record that cannot be regenerated.

The fix

They have been durable since 2026-07-29 and nothing pointed at them:

/gscratch/makelab/jonf/rampnet_yolo_baseline_51 — klone lab storage, outside the
~21-day /gscratch/scrubbed purge that this record exists to survive. 1.8 GB; each arm
directory holds best.pt, last.pt, args.yaml, results.csv. A MANIFEST.md there
already carries the snapshot time, the job↔config map, per-arm epoch/mAP state, and
sha256 for all 10 .pt files verified against source at copy time — it simply was not
referenced from the record.

Two caveats included rather than left implicit

• It is a snapshot, not a live mirror, taken 2026-07-29T20:51Z while all five arms were
  still training. They have advanced since — y11l_pano 13 → 15 epochs, y26_pano 18 → 20
  as of 2026-07-30 — so it must be re-taken before anything is quoted as final, and
  /gscratch/scrubbed/jfroehli/yolo_runs/<arm>/results.csv stays the live source until then.
• y26_tiles_l40s is missing entirely. That arm postdates the snapshot (Slurm 37889646,
  forked to a dedicated gpu-l40s node on 2026-07-29 to escape the ckpt preemption ceiling)
  and needs adding on the next pass.

Staging to Hugging Face is still worth doing for external reproducibility, and the section
says so — it is just no longer the only thing standing between these weights and the purge.

Related: #51 · #91 (Tillicum onboarding, which re-snapshots as part of the migration)

🤖 Generated with Claude Code (claude-opus-5[1m])

[jonfroehlich]

Closing as redundant — PR #87 already fixes this TODO, and more thoroughly than this branch does. I flagged it as an open gap after reading main, without checking that an open PR already closed it. #87 additionally documents the sha256 verification result (10 ok / 0 mismatched), the /gscratch/scrubbed purge rationale, and why last.pt is kept alongside best.pt (optimizer/EMA state for resume — exactly what the y26_tiles_l40s fork needed). Both would have edited the same lines and conflicted.

Two small deltas from this branch that #87 does not currently carry, noted here so they are not lost:

1. The snapshot has measurably drifted. Refresh the YOLO baseline record: durable weights, ckpt ceiling, fork (#51) #87 says it was staged "while the runs were still going"; the concrete drift as of 2026-07-30 is y11l_pano 13 → 15 epochs and y26_pano 18 → 20. Worth an explicit "re-snapshot before quoting as final," with /gscratch/scrubbed/jfroehli/yolo_runs/<arm>/results.csv named as the live source until then.
2. y26_tiles_l40s's own weights are not in the durable snapshot. Refresh the YOLO baseline record: durable weights, ckpt ceiling, fork (#51) #87 mentions the fork as a consumer of last.pt, but the fork postdates the 2026-07-29T20:51Z copy, so its weights are not covered by it.

Happy to add both to #87 as a commit if wanted.