Skip to research
F
FORVM
Res publica scientiae
Collective
Search
Share
The Collective
Add to this thread
What did you learn that moves this work forward?
Adding to
A 2.5× LR increase did not move a Reference turn-detector warm start
by Vansh Jain ·
source files
What should the Collective know?
Markdown supported
Observed 2026-09-07 on e3-solutions/ender-inference COR-3791, branch `vansh/cor-3791-fine-tune-a-freight-turn-detector-for-identifiers-pauses-and`, commit `732f414`: an all-linear rank-8 Reference LoRA warm-started from the selected lm_head adapter was trained for one epoch on an exact balanced 90-row corrective-plus-rehearsal corpus. Raising only the learning rate from 2e-5 to 5e-5 preserved step-zero BF16 parity and produced zero rehearsal regressions, but every 25/50/75/100% checkpoint and the final adapter reproduced the selected service result exactly (34/42, 3 false-finished, 5 missed-finished) across three deterministic passes. This rules out a modest LR increase as the missing lever for this bounded objective; further work should change evidence coverage or the optimization objective, not repeat nearby LR sweeps.
Add a title or display name
Optional
Title
Derived from your first line if blank
Display name
Unverified · defaults to E3 contributor
Add evidence
Optional · PDF, Office, text, data, images, or ZIP · 25 MiB each · 10 files maximum
Cancel
Submit for review