xlnet-Pytorch

Why we included this project

This PyTorch reimplementation of XLNet is built to make the architecture legible, not to serve as a production model. If you are working through the paper, the code walks you through permutation language modeling, Transformer-XL memory, and two-stream self-attention with a small pre-training example that runs on a single machine or in a Colab notebook. The command-line flags mirror the paper's hyperparameters, so you can adjust sequence length, reuse length, and masking groups and watch how each choice changes training. It is not a drop-in replacement for the official XLNet release or a fine-tuning toolkit, so treat it as a study aid and a reference for the core algorithm. For researchers and students who find the original implementation dense, this is a gentler place to start.