Skip to content

test(classifier): update test harness for classifiers - #454

Open
tachyonicClock wants to merge 4 commits into
adaptive-machine-learning:mainfrom
tachyonicClock:test_harness
Open

tachyonicClock wants to merge 4 commits into
adaptive-machine-learning:mainfrom
tachyonicClock:test_harness

Conversation

@tachyonicClock

@tachyonicClock tachyonicClock commented Oct 7, 2026 •

Copy link
Copy Markdown
Collaborator
  • Adds a test_classifier suite of tests parameterized by classifier.yml.
  • Splits one big test per method into many focused small tests per method.
  • New determinism test and adds TODOs for identified determinism issues.
  • Required some changes to how Finetune the pytorch based method is built.

Closes: https://github.com/adaptive-machine-learning/backlog/issues/161

@tachyonicClock
tachyonicClock force-pushed the test_harness branch 2 times, most recently from 0a6a567 to ffc0980 Compare October 7, 2026 02:50
Comment thread docs/contributing/faq.md
@tachyonicClock tachyonicClock changed the title test(classifier): update test harness for classifiers WIP: test(classifier): update test harness for classifiers Oct 7, 2026
tachyonicClock and others added 3 commits October 8, 2026 11:38
CI measured 89.2% cumulative accuracy for ShrubsClassifier on
ElectricityTiny vs. the 89.15% baked into classifier.yml, a gap equal to
one flipped prediction out of ~2000. ShrubsClassifier trains via
per-instance SGD over many sequential floating-point updates, so a
sub-ULP difference between platforms/BLAS builds can compound and flip a
single borderline prediction. Determinism, predict, and save/load tests
for this case are unaffected, so this isn't a logic bug.

Add an optional accuracy_abs_tolerance field to the test harness Case,
defaulting to None (unchanged exact comparison for all other cases), and
set it to 1.0 for ShrubsClassifier.

Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Fix grammar, a copy-pasted classifier name, a contradictory abstention
note, and trim an over-long CSMOTE explanation.

Assisted-by: claude-code:claude-sonnet-5
…P drift

Assisted-by: claude-code:claude-sonnet-5.5
@tachyonicClock tachyonicClock changed the title WIP: test(classifier): update test harness for classifiers test(classifier): update test harness for classifiers Oct 7, 2026

This branch has not been deployed

No deployments
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