all_results.json

{
    "test_BLEURT": -1.4601,
    "test_WER": 0.5257,
    "test_loss": 1.3316419124603271,
    "test_rouge1": 0.644,
    "test_rouge2": 0.3634,
    "test_rougeL": 0.5779,
    "test_rougeLsum": 0.5777,
    "test_runtime": 31.573,
    "test_samples_per_second": 11.941,
    "test_steps_per_second": 1.995
}