Skip to content

fix(tasks): order readout probabilities by the task's labels - #212

Merged
KartikP merged 1 commit into
unified-model-interface-v2from
fix/readout-probability-order
Nov 4, 2026
Merged

KartikP merged 1 commit into
unified-model-interface-v2from
fix/readout-probability-order

Conversation

@KartikP

@KartikP KartikP commented Nov 4, 2026

Copy link
Copy Markdown
Contributor

Problem

With a readout, behavior:choice listed its columns in the task's label order (Choice(['animal', 'vehicle'])), but behavior:probabilities listed them in the readout classifier's own order (here vehicle, animal). The coordinates were right, so .sel(choice='animal') worked, but .values of the two assemblies disagreed column for column. A rebuilt lesson computed p(correct) = 0.21 at 83% accuracy because of it.

Fix

Readout probabilities are reordered to Choice.labels, on both the readout path and next to a generated answer. Likelihood probabilities already followed the labels.

Tests

  • New: test_readout_probabilities_follow_the_task_label_order, for both label orders. The reversed order fails without the fix.
  • Core suite: 1,190 passed. The 7 errors and 1 failure are local environment tests (submission database, conda_score.sh), unrelated.

behavior:choice followed Choice.labels but behavior:probabilities
followed the readout classifier's own label order, so the two
assemblies' .values disagreed column for column.
@KartikP
KartikP merged commit 5193fe7 into unified-model-interface-v2 Nov 4, 2026
8 checks passed
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