cast_checkpoint/optimizer_spec.json

{"__ctor": "palivla.optimizer.make_optimizer", "config": {"base_learning_rate": 0.0001, "num_train_steps": 200000, "optimizer": "adamw"}}