Plug-and-Play LM Checkpoints with TensorFlow Model Garden

Install tf-models-official, pick a model (BERT/ALBERT/ELECTRA), download the checkpoint, and construct an encoder via EncoderConfig from either params.yaml (new) or legacy *_config.json. Wrap the encoder with tfm.nlp.models.BertClassifier for a 2-class head, then restore only encoder weights with tf.train.Checkpoint(...).read(...) (the head stays randomly initialized). For ELECTRA, discard the generator and use the discriminator (encoder) for downstream tasks. This gives a ready-to-fine-tune classifier across the BERT family with minimal code.