{
    "eval_reasoning_loss": 0.661374032497406,
    "eval_reasoning_runtime": 89.2902,
    "eval_reasoning_samples_per_second": 1.445,
    "eval_reasoning_steps_per_second": 0.37,
    "eval_samples": 100,
    "eval_utility_loss": 0.9463363885879517,
    "eval_utility_runtime": 33.9331,
    "eval_utility_samples_per_second": 1.444,
    "eval_utility_steps_per_second": 0.383
}