Chinese-XLNet

Why we included this project

Chinese-XLNet offers pre-trained checkpoints of the permutation-based XLNet model, which captures bidirectional context through a different mechanism than the masked BERT-style encoders most Chinese NLP projects build on. It ships two sizes, a 12-layer base model and a 24-layer mid variant with over 200 million parameters, both trained on the same 5.4-billion-token corpus used for HFL's well-known Chinese BERT releases. You can load the weights through Hugging Face Transformers in a couple of lines and fine-tune them for extractive reading comprehension, sentiment classification, or other sequence tasks. The README documents the pretraining details, fine-tuning settings, and baseline results against Chinese BERT, so the repo works as both a practical starting point and a useful reference when comparing encoder choices.