Skip to content

calibrated 2.0.1

Latest
Latest

Choose a tag to compare

View all tags
@m-sanchez m-sanchez released this 03 Oct 08:57
v2.0.1
5ebef32
This commit was created on GitHub.com and signed with GitHub’s verified signature.
GPG key ID: B5690EEEBB952194
Verified
Learn about vigilant mode.

Numerical corrections for temperature scaling. npm registry publication is pending; the corrected source is available in PR #1.

  • Compute NLL without probability clipping and preserve raw-logit decisions when probabilities round to ties.
  • Handle extreme finite logits, validate inputs and search settings, and report fit status and iteration count.
  • Evaluate the seeded demo on separate calibration and test data; clarify the limits of ECE resampling and null simulations.
  • Compatibility: empty NLL returns NaN; empty fitting and empty softmax throw. Invalid temperatures, dimensions and search settings throw. Code constructing TemperatureFit objects must supply the new status and iterations fields.
  • Validation: 46 tests, typecheck, build and package-install checks passed on Node 22, 24 and 26 in CI. Independent local SciPy checks passed for 403 probability/loss fixtures and three fits.

Detailed changes and limits.

Assets 3
Loading