Skip to content

Weight OTB general and math splits independently - #126

Open
rudycelekli wants to merge 1 commit into
facebookresearch:mainfrom
rudycelekli:fix/otb-disjoint-macro-splits
Open

rudycelekli wants to merge 1 commit into
facebookresearch:mainfrom
rudycelekli:fix/otb-disjoint-macro-splits

Conversation

@rudycelekli

Copy link
Copy Markdown

format_metrics includes math rows in the general mean, then averages that mean with the math mean again. This double-weights math and makes the intended macro average depend on relative split sizes. A fixture with general accuracy/score 1 and math 0 reports 0.25 instead of 0.5; token means are similarly skewed.

Exclude math rows from the general masks for both overthinking and underthinking. The regression checks accuracy, score, token means, F1, and invariance when identical general rows are duplicated. It fails on main and passes with this change.

Validation: Python 3.9.6 and real pandas, python3 -m unittest discover -s projects/otb/tests -v (1 regression passed on this branch). Model imports are isolated; no model/provider calls, GPU benchmarking, full-suite run, or hosted CI result is claimed. Based on main 6bcc4813bfb72ec1dcec7923e449c027a4d9194e.

@meta-cla meta-cla Bot added the CLA Signed This label is managed by the Meta Open Source bot. label Nov 4, 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

CLA Signed This label is managed by the Meta Open Source bot.

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant