{
    "eval_reasoning_loss": 0.9136209487915039,
    "eval_reasoning_runtime": 29.1446,
    "eval_reasoning_samples_per_second": 4.426,
    "eval_reasoning_steps_per_second": 1.132,
    "eval_samples": 100,
    "eval_utility_loss": 1.1523499488830566,
    "eval_utility_runtime": 11.0014,
    "eval_utility_samples_per_second": 4.454,
    "eval_utility_steps_per_second": 1.182
}