{
    "train_batch_size": 128,
    "gradient_accumulation_steps": 1,
    "optimizer": {
        "type": "Adam",
        "params": {
            "lr": 0.00015
        }
    },
    "zero_optimization": {
        "stage": 2
    }
}