v0.3.0

Added

• Add "ridge_probe" finetuning strategy: closed-form streaming ridge regression probing on frozen backbone features. Single pass over the dataloader, eigendecomposition-based λ sweep, no hyperparameter tuning needed. For high-dimensional backbones, max_features (default 5000) triggers a Gaussian random projection seeded by Experiment.seed; multi-seed runs therefore produce different projections and enable variance estimation (#21).
• Add FlattenHead head type (used internally by ridge probing) (#21).
• Add RidgeProbingTraining config with kind="ridge" discriminator for the Training union (#21).
• Add pytest to the CI workflow to run tests on each pull request (#18).
• Add max_meta_experiments argument to helpers.run_multiple_per_node() as alternative to max_experiments_per_node (#17).
• Add training_required_modules parameter to the backbones (#20).
• Expose the preload parameter of datasets (#26).
• Add CITATION.cff metadata file and a Zenodo DOI badge in the README for archival citation via Zenodo (#27).

Changed

• Popularize the use of import open_eeg_bench as oeb via the README and documentation (#17).
• Improve import hints via the __all__ variable in __init__.py (#17).
• All backbones now use peft_target_modules="all-linear" by default for simplicity, which leads to a slight increase in the number of parameters being finetuned (#20).
• Change default LoRA.bias to "none" to match PEFT's default (#20).
• Allow disabling dropout layers of the backbone (default: True) (#20).
• Change default Experiment.seed from 42 to 0 (#21).

Fixed

• The lazy modules are now initialized before and after applying the PEFT (necessary for some PEFT methods like OFT) (#17).
• Take into account BENDR's channels projection fix in the BENDR backbone (#20).

Full Changelog: v0.2.1...v0.3.0