SeanLee97/AnglE
A library for training and inferring state-of-the-art sentence embeddings using BERT and LLM backbones, optimized for dense retrieval and RAG applications.
Velocity · 7d
+0.6
★ / day
Trend
→steady
star history
AnglE provides implementations of various embedding loss functions including AnglE loss (ACL24), Contrastive loss, CoSENT, and Espresso loss (ICLR 2025). It supports training on BERT-based models (BERT, RoBERTa, ModernBERT) and LLM-based models (LLaMA, Mistral, Qwen, OpenELMo). The library enables single and multi-GPU training for creating powerful text embeddings used in semantic search, dense retrieval, and retrieval-augmented generation pipelines.