lightonai/pylate

Builds on Sentence Transformers to simplify ColBERT fine-tuning with contrastive loss and knowledge distillation, supporting single and multi-GPU training with memory-efficient techniques like GradCache. Enables constructing ColBERT models from arbitrary pre-trained language models and streamlines dense retrieval through flexible document indexing and ranking. Integrates with Hugging Face datasets and provides built-in evaluators including NanoBEIR for rapid validation on retrieval benchmarks.

742 stars. Used by 3 other packages. Actively maintained with 2 commits in the last 30 days. Available on PyPI.