Repository navigation
Prepare neural run 3: min epochs, held-out selection split - #40
Merged
Merged
Conversation
TrainConfig.min_epochs keeps early stopping from ending a run before it starts predicting (run 2's uniform-weight runs stopped at epoch 4). corpus split-dev draws splits/selection-v1 from the model-drafted dev batches by leakage group: 111 documents for early stopping and run selection, 61 held out for comparisons. train/experiment take --dev-subset and evaluate/compare take --subset. Grids distilbert-3 and deberta-3 run up to 20 epochs with min_epochs 6.
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
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
Summary
TrainConfig.min_epochs: early stopping may end a run only after this many epochs. In run 2, 11 of the 12 uniform-weight runs scored 0 at epoch 1 and were stopped at epoch 4, before they predicted anything. New test (parametrised) confirms a constant-score run trains formin_epochs.triggermap corpus split-devdraws an immutable selection / held-out split of dev gold by leakage group (seed 0).splits/selection-v1/holds the 172 model-drafted dev records ofbatch-03+batch-06:The manifest records the gold file hashes, the split hash and the list hashes.
--dev-subset(train,experiment) and--subset(evaluate,compare) take a JSON list of advisory ids. Tests cover both andsplit-dev.distilbert-3/deberta-3: same 24-run grid, up to 20 epochs,min_epochs = 6, patience 3.run_neural.shstops early on the selection half;score_neural.shscores and compares on the held-out half and on the 10 hand-labelled documents. Smoke-tested on the run-2 archive.docs/training_on_gpu.mdupdated.This removes run 2's caveat: comparisons on the held-out half are not biased by run selection. The test split is not touched.
Checks
ruff check,ruff format --check,mypy src/,pytest -m "not slow"(241 passed);shellcheckclean; both grids dry-run to 12 runs.