PCNtoolkit version 1.3.0 is out!

Announcements
1

You can update by running pip install --upgrade pcntoolkit

Here are the highlights:

  • Improved evaluation metrics (MLL and MACE) - and there’s now a tutorial on the website walking through what each metric actually does

  • You can now load pre-trained models saved with older PCNtoolkit versions and use them in the newest pcntoolkit version

  • More detailed, user-friendly tutorials on model merging and federated learning

  • Better contribution guidelines and GitHub standards

  • Bug fixes, bug fixes, bug fixes :bug:

Thanks to everyone who contributed! And looking forward to more contributions!

Full changelog

Major changes:

  1. Add fractional polynomials basis function

  2. Rename NLL (negative log-likelihood) → MLL (mean log-loss)

  3. Add MACE averaging (as defined in Zamanzadeh et al. (2026))

  4. Add kurtosis and skewness as evaluation metrics

  5. Add migration strategy for saved models

  6. Make HBR SHASHb faster by adjusting the dp parameter

  7. Manage matplotlib more flexibly

  8. Refactor BLR tutorial

  9. Add federated learning and evaluation metrics tutorials

  10. Update merge tutorial

Minor changes:

  1. Update contributing guidelines and add rules in GitHub (Issue and PR templates)

Bug fixes:

  1. Add upper bounds for pymc, nutpie and pytensor

  2. Fix HBR SHASH to return mean and second moment in m1m2()

  3. Add h5py and h5netcdf as dependency in toml