Skip to content

results: learning-rate pilot selects 5e-5 for both encoders - #9

Merged
drewOrc merged 2 commits into
mainfrom
results/pilot-lr
Sep 23, 2026
Merged

drewOrc merged 2 commits into
mainfrom
results/pilot-lr

Conversation

@drewOrc

@drewOrc drewOrc commented Sep 23, 2026

Copy link
Copy Markdown
Owner

Validation-only LR pilot (k=100, seed 42). Selected 5e-5 for BERT (reused AC2 run) and ModernBERT. Both optima are at the top of the frozen grid {1e-5, 2e-5, 5e-5}; the grid is not widened after seeing results. Fills configs/curve.yaml learning rates and adds a DEVLOG entry.

Validation only, k=100, seed 42. BERT 5e-5 reuses the equivalent AC2
run. Both optima sit at the top of the frozen grid; the protocol is not
widened after seeing the results, and the README will state the limit.
The old test pinned the protocol to its unfilled state and broke as soon
as the first pilot value was copied in. The new test compares every
value in configs/curve.yaml with the selection recorded in
results/pilots/*.json, and requires null while a pilot output is absent,
so a mistyped value fails CI.
@drewOrc
drewOrc merged commit 7c7b166 into main Sep 23, 2026
4 checks passed
@drewOrc
drewOrc deleted the results/pilot-lr branch September 23, 2026 09:05
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