finetune

Why we included this project

This library wraps TensorFlow implementations of BERT, RoBERTa, GPT, GPT-2, DistilBERT, and a couple of convolutional sequence models behind a scikit-learn style API. Instead of writing a training loop, you instantiate a Classifier, call fit() on your data, and the library handles finetuning, saving, and reloading. It also supports a two-stage workflow: finetune on a large unlabeled corpus first, then continue on a smaller labeled set, which helps when labeled data is thin on the ground. The Docker setup keeps GPU and CPU runs straightforward, and the estimator interface covers classification, regression, and sequence labeling. One honest caveat: it targets the TensorFlow 1.x era.