eval_results.json

{
    "epoch": 90.0,
    "eval_log_ratio_diff/max": -44.0,
    "eval_log_ratio_diff/mean": -44.0,
    "eval_log_ratio_diff/min": -44.0,
    "eval_log_ratio_diff/std": NaN,
    "eval_logits/chosen": -1.75,
    "eval_logits/rejected": -1.75,
    "eval_logps/chosen": -494.0,
    "eval_logps/rejected": -426.0,
    "eval_loss": 17.31753921508789,
    "eval_nll_loss": 1.65625,
    "eval_ref_probs/chosen": 0.0038271306548267603,
    "eval_ref_probs/rejected": 0.0034569373819977045,
    "eval_rejected_term/max": 0.0011350036365911365,
    "eval_rejected_term/min": 0.0011350036365911365,
    "eval_rewards/accuracies": 0.3400000035762787,
    "eval_rewards/chosen": -17.25,
    "eval_rewards/margins": -4.40625,
    "eval_rewards/rejected": -12.8125,
    "eval_runtime": 25.6175,
    "eval_samples_per_second": 39.036,
    "eval_steps_per_second": 9.759
}