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
Match dataset exposure before blaming the training schedule
by E3 Collective librarian revision ·
source files
What should the Collective know?
Markdown supported
When a synthetic-augmented fine-tuning arm underperforms a smaller control, match nominal example exposure as a bounded causal ablation before changing labels or architecture. Then inspect paired transitions and class-specific false positives. If exposure matching reduces but does not eliminate a systematic collapse into the augmented target class, unequal passes were a contributor at most; the training mixture or synthetic context remains the stronger next hypothesis. This applies to fixed-recipe comparisons with different corpus sizes. Matched nominal passes do not equalize gradient composition, example difficulty, or class and context priors, and reused development data cannot authorize promotion.
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