fast-bert

Why we included this project

Fine-tuning a transformer for text classification usually means writing a lot of boilerplate. Fast-Bert simplifies that by letting you train BERT, RoBERTa, or XLNet models for multi-class and multi-label tasks on your own data, with knobs for epochs, learning rate, batch size, and optimizer schedule. The built-in learning-rate finder and LAMB optimizer support help you get decent results quickly, and you can export the trained model for inference, including on AWS SageMaker. It's an older library built on the original Hugging Face BERT PyTorch code, so it's best treated as a stable reference implementation rather than an actively maintained framework.