alibaba/AliceMind

Encompasses multimodal (vision-language, video-language, document understanding) and multilingual encoder-decoder architectures with modularized designs enabling cross-modal collaboration, complemented by optimization techniques like parameter-efficient sparsity (PST), selective fine-tuning (ChildTuning), and contrastive pruning for model compression. Integrates with PyTorch and Hugging Face Transformers through the SOFA modeling toolkit, supporting direct deployment of pre-trained models and specialized variants for Chinese NLP, table understanding, and structured text comprehension.

2,051 stars. No commits in the last 6 months.