Test53 training modes

What I was poking at

Experiment 01 asked whether cameral width and learning rate mattered under Lucy live-fit. Experiment 02 turns the dial the other way: fix the dayroute harness and sweep the whole Welvet training-mode matrix. Test53 is the “dense × mode × dtype × funny-LR” farm — can the net learn while it still serves, across all 29 named train modes, lots of storage types, and learning rates that classical SGD would never touch.

The overview pack is the Welvet 1.0.3 training-modes writeup, a short cloud-free edge-learning audio briefing, and an Ocean compare that lines up the same recipes across cam1 vs cam3 over eight LR steps (0.02 → 100m). Under that sit four Mac run folders with Lucy PDFs plus the full history / LPD / progress / results JSON.

The setup, in plain terms

Each Lucy report is a Tide dayroute snapshot. Lucy Score is still Throughput × Availability × Accuracy / 10,000. Modes follow the same legend as the cameral sweeps: [T]=Tween, [S]=Split, [FP]=FastProxy, [L]=Linear, [HP]=HeadProxy. Classical BP (NormalBP / StepBP / MeshBP / plain SGD) locks the serve path and burns Availability. Tween and Split families keep inference alive while they update.

The M4 “low” folders are not a single polite rate — they finish 3,944 cells spanning lr 0.02, 2, 200, and 2k, once as single-cam dense and once as tricameral dense. The M5 “high” folders push funny rates hard: the dense Lucy PDFs centre on lr 20k (and the cam3 JSON also walks 1m / 10m / 100m). The M5 high JSON also grew into a wider unfinished multi-layer tree (about 10.6k of 63k cells) with an even hotter mid-run leader — the PDF remains the dense dayroute board for that host.

What showed up

Tween wins the farm. On the finished M4 single-cam matrix, Tween-family cells averaged roughly 731 Lucy Score versus about 394 for BP-family cells, and the absolute champion was dense|float64|MeshTween at lr 2k (score 3968). Tricameral M4 finished with MeshTweenSplitSparse float32 at lr 0.02 on top (3777) — Split/Sparse keeping Availability high (~43%) mattered more than chasing the wildest rate.

On M5 dense high, the Lucy board crowned MeshTween uint8 at lr 20k (4058). With cameral-3, MeshTween float32 at lr 20k led (3477), while Availability champions pushed into the seventies with TweenSplitSparse. Ocean’s cam1-vs-cam3 compare is quieter on accuracy: cam1 edged mean Acc by a fraction of a point in both HI and LO bands. The interesting gap is still train-mode family and LR band, not “more cams automatically smarter.”

LPD / gold-standard rows keep pointing at tiny discrete cells that still keep Acc. Gold-ish modes cluster around Step/Tween Split Sparse and FastProxy sandwiches — small RAM, real live-fit pillars — while binary keeps lighting the mobile-score trap. The training-modes research note says the quiet part out loud: on storage-truth discrete weights, mid-to-high learning rates are how Tween actually bit-flips the microkernel, not a mistake to avoid.

What I take from it

If Experiment 01 said cameral width and LR are a joint surface, Experiment 02 says the train-mode family is the bigger fork in the road. Classical backprop is the wrong tool for continuous serve-while-learn on this engine. Mesh/Step Tween with Split, Sparse, and proxy variants are where Lucy Score and Availability live; funny LRs belong with them once you leave Float32-master thinking behind.

Read the overview PDF/MP3 for the mode taxonomy, the Ocean compare for cam1 vs cam3 LR curves, then the four run bundles for the raw boards. The champion table below uses each dense Lucy report’s Lucy Score winner; JSON best_ids match those finished dense farms (and show the wider M5 multi-layer mid-run separately in progress.json).