LibAUC 2.0.0

New Features

• Added the LibAUC Trainer, enabling users to easily set up and run training pipelines.
• Added distributed training support for all optimizers used with AUCMLoss, AveragePrecisionLoss, pAUC_CVaR_Loss, pAUC_KL_Loss, tpAUC_CVaR_Loss, tpAUC_KL_Loss, ListwiseCELoss, and NDCGLoss.
• Integrated Hugging Face models into the LibAUC framework.
• Updated AUCMLoss and PESG to support resume training. All optimizers now support checkpoint-based training resumption.

New Optimizers

• Added two algorithms for Extreme Classification: SCENT and SOX, as introduced in our recent work: https://arxiv.org/abs/2602.02877
• Added the STACO algorithm for optimizing two-way partial AUC: https://openreview.net/pdf?id=M3kibBFP4q

New Tutorials

• Added a tutorial on optimizing global contrastive loss with iSogCLR for unimodal contrastive learning: https://docs.libauc.org/examples/isogclr_unimodal.html
• Added a tutorial on training linear models for extreme classification using SCENT: https://docs.libauc.org/examples/scent_extreme_classification.html
• Added a tutorial on training CLIP models using FastCLIP-v3: https://docs.libauc.org/examples/fastclip.html
• Added a tutorial on distributed optimization of various losses using multiple GPUs: https://docs.libauc.org/examples/dist_train.html
• Added a tutorial on using the LibAUC Trainer: https://docs.libauc.org/examples/libauc_trainer.html

Bug Fixes

• Fixed a tensor shape mismatch issue when training models with iSogCLR.
• Fixed a sampler bug triggered under extremely small imbalance ratios.

Key contributors

• Siqi Guo, Gang Li, Xiyuan Wei, Linli Zhou, Xingyu Chen, Tianbao Yang