Phase 2.1A step budget v2

Date: 2026-08-27

Decision

Replace the inherited 15,000-step fixed run with a 3,000-step hard cap and pre-registered ID-dev early stopping. Apply the same cap and selection rule to Robometer-Low-LoRA and the future full-decoder Robometer-Low comparison.

Scale calculation

Frozen train set: 2,021 preference pairs.
Effective batch: 64 pairs per optimizer step.
15,000 steps: 960,000 pair presentations, approximately 475 nominal passes.
3,000-step cap: 192,000 pair presentations, approximately 95 nominal passes.
Minimum 1,000 steps: 64,000 pair presentations, approximately 32 nominal passes.

The official 15,000-step value was copied from the million-scale RBM-1M recipe; it was not estimated from this 2,021-pair Low split.

Evidence from the stopped v1 run

The user-stopped seed-7 run logged through step 246. Mean total train loss fell from 4.573 over steps 1-25 to 3.218 over steps 201-225; the last-100-step linear trend was still negative by about 0.455 per 100 steps. It had not converged. However, the 15,000-step scheduler used 750 warmup steps, so step 246 had reached only about 33% of the target learning rate. This curve cannot be extrapolated to justify 15,000 steps.

Selection protocol

Scheduler: cosine with 5% warmup; 150 warmup steps at the cap.
Evaluate every 250 steps on the original 108-pair dev split.
Evaluate both A/B orientations for each pair.
Select by the sum of mean preference, progress, and success ID-dev losses.
Minimum training: 1,000 steps.
Early stop after four consecutive evaluations without an improvement of at least 0.005 ID-dev loss.
Restore the best ID-dev checkpoint for export.
Do not use confirmation_dev or RBM-EVAL-OOD for selection.

The cap is a safety bound, not a claim that every seed needs exactly 3,000 steps. The selected step is an experimental result recorded per seed.

Restart boundary

The v1 run is not resumed into v2. Its latest complete checkpoint is step 4, and v2 changes both the scheduler horizon and checkpoint-selection protocol. All v2 seeds must start from the same pinned base model in a new output root.