Freeze approved H2 S1 fold0 last-epoch200 training amendment

This commit is contained in:
Frank14f committed 2026-11-17 16:08:56 +08:00
1 parent a864b29563
commit 7a7174746d
8 files changed
+345

No files matched your search

+11
View File
@@ -0,0 +1,11 @@
# H2-S1 labelled engineering amendment
Claude1794901608.131369 supersedes the pilot's 16 GPU-hour cost gate with staged 40 GPU-hours including the original pilot. Original contract9f616974, PARKED_RESOURCE and all pilot artifacts remain unchanged. This is a resource decision before any formal H2 result, not a model-selection decision.
S1 is one fresh seed0/fold0 fit: original19 features, FP64 F/c skip, full original nodes and edges, width64/four blocks, exact original optimizer and 200 epochs, ten-epoch warmup/cosine. Reuse only learning-side normalization and original learning c, never pilot weights. Validation every10 epochs is diagnostic only; last epoch200 is the sole checkpoint. No early stopping or best-checkpoint selection.
Unique root artifacts/phase2-h2-s1-2026-11-17; single V100 GPU0, workerCPU0–1, guardCPU2. Training cap25200 GPU-seconds includes60 setup and all waits, failures and monitoring; CPU cap25200 seconds includes60 reserve. Original pilot182.352516340965 GPU-seconds is counted once. Prediction allowance<=600 GPU-seconds and CPU postfit<=3600 seconds will be enforced separately before execution, with all GPU charges included in total144000. This initial entry starts training only; it cannot launch predictions, scoring or S2. No retry. Allocated/reserved15GiB, whole treeRSS16GiB, disk20GiB; cumulative JSON history/progress/failure/result peak guard plus device sampling, limits sampled and any final overshoot fails.
After train PASS and exact epoch200 checkpoint freeze, freeze test predictions and scoring source before reading test targets. Compare with the original A on the same fold0 test components and query metrics. S2 folds1–4 are eligible only if paired component field-L2 H2−A CI upper<+0.005 OR last-fill-distance H2−A CI upper<0. Strict bounds, original2000 paired bootstrap/seed, no undefined-bound pass. S2 remains gated, not running. CPU field scoring and topology costs are distinct; this S1 gate does not require new topology computation.
Seven CPU engineering regressions passed: strict OR threshold, final-only checkpoint, reject validation normalization, schedule indexing, historical failure peak retention, no test/pilot-weight read in train, unapproved-scope rejection. Existing model gradient/data acceptance remains unchanged and is not rerun. These are CPU checks, not evidence of actual S1 training. Source closure and inherited inputs are pinned before launch; actual started/worker progress alone establishes execution.