Skip to content

bench: larger synthetic n_train + L=4 Stage-2 recast#113

Draft
davorrunje wants to merge 2 commits into
mainfrom
chore/synth-larger-for-large-batch-band
Draft

bench: larger synthetic n_train + L=4 Stage-2 recast#113
davorrunje wants to merge 2 commits into
mainfrom
chore/synth-larger-for-large-batch-band

Conversation

@davorrunje

Copy link
Copy Markdown
Owner

What

Two Stage-2 synthetic-benchmark tweaks:

  • bump synthetic n_train 16000 → 32000 (crosses the 20k large-batch band; more data for the deep arms)
  • recast the Stage-2 depth analysis around the L = 4 sufficiency threshold

Draft for tracking — this branch had no prior PR.

🤖 Generated with Claude Code

davorrunje and others added 2 commits July 13, 2026 11:31
…re data for deep)

At 16000 rows the synthetic datasets fell in the small-batch band ([8..256]),
so an Optuna trial that drew batch=8 ran ~2000 steps/epoch (300k tiny GPU
launches/trial at ~9% util) — the full synthetic sweep was ~days. Bumping to
32000 (>20k) puts them in the large-batch band ([512..4096]) => ~10 steps/epoch,
~100x faster, AND gives the deep arm more data for a fairer depth signal.
Non-degeneracy R2<0.7 gate + relu≠elu + monotonicity all still hold at 32k.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
…hold

Recasts the Stage-2 results on the effective-monotone-layer axis L (plain:
depth+1; residual/deep: 2*depth+2) rather than the plain/residual/deep flavor
split, per the depth-4 monotone universal-approximation theorem (4 layers
suffice). Reports best-at-L<=4 (sufficient) vs best-at-L>4 (excess) per dataset
in its primary metric, plus an error-vs-L curve from the Optuna search trials
(best CV objective at each sampled L, L=4 line marked). Reads committed results;
no training.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant