v1.4.0

31.03.2026 - Version 1.4.0

• Bioengineer: Adding new bioengineer module with base class, baselines, and tokenizer refactoring
• ProteinGym: Added support for ProteinGym datasets (SC/NDCG calculation) in BioEngineer
• Feature Scaling: Introduced automatic feature scaling for inference methods and added support for feature scaling
  in PBC tasks with standard and minmax scaling methods for per-sequence and per-residue embeddings
• Added new metrics NDCG and balanced/per-class accuracy
• AutoEval Enhancements:
  • Revamped documentation and added a dashboard for better visualization
  • Implemented new features such as report publishing, comparison of reports, and enhanced dashboard views
  • Added frontend support for loss curve visualization with indicators for the best epoch and improved ranking
    utilities
  • Fixed adaptation of updated PBC datasets with proper configuration handling
• Refactoring and Fixes:
  • Adapted inference methods for new prediction types and structures with feature scaling support
  • Improved early stopping logic with a relative threshold method and updated related configurations
  • Enhanced tokenizer compatibility with Transformers v5 API
  • Various fixes, including handling deprecated tasks, improving plotting, and resolving minor bugs in the autoeval
    pipeline
• Added new examples, such as plm_eval_zeroshot, and improved PBC task handling with updated configuration options
• Refactored multiple components for better performance, consistency, and readability
• Active Learning: Added Gaussian Processes (AL) support (wip)
• Uncertainty Quantification: Added uncertainty to prediction sets for improved model confidence estimation