VinAIResearch/PhoBERT

Provides base and large transformer models (135M–370M parameters) optimized via RoBERTa's pre-training procedure and trained on 20GB of Vietnamese Wikipedia/news corpora. Integrates seamlessly with Hugging Face `transformers` and `fairseq` frameworks, with models available on the Hub; requires upstream word segmentation via VnCoreNLP's RDRSegmenter to handle Vietnamese morphology before inference. Achieves state-of-the-art on four downstream tasks: POS tagging, dependency parsing, NER, and natural language inference.

775 stars. No commits in the last 6 months.