Speed up training and add OOD deferral analysis#5
Merged
Conversation
sampling_strategy="unique" generated ~1.2M contrastive pairs on the grown pool (~60 min runs that kept dying), most of them trivially-easy near-duplicate pairs from the disfluency augmentation. oversampling with num_iterations=40 draws ~400 pairs, trains in ~2 min, holdout accuracy 0.925.
report.py scores an OOD/garbled probe set next to the holdout and adds two figures: a confidence histogram and a deferral-tradeoff curve. They show routelet is overconfident (max-softmax ~0.98 even on gibberish), so the 0.55 gate deferred almost nothing; ~0.95 is the usable operating point.
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Two changes. (1) Training: swap the exhaustive 'unique' pair sampling for oversampling (num_iterations=40, 1 epoch), ~60 min runs that kept dying become ~2 min; holdout accuracy 0.925. (2) Report: score an OOD/garbled probe set next to the holdout and add a confidence histogram + deferral-tradeoff curve. They show routelet's max-softmax saturates near 0.98 even on gibberish, so the 0.55 gate deferred ~0% of OOD; ~0.95 is the usable operating point (paired with the matching aegis tuning change).