eval_results.json

{
    "epoch": 2.999098751126561,
    "eval_kl/chosen": 0.00024373203632421792,
    "eval_kl/rejected": 0.00031893892446532845,
    "eval_logits/chosen": -1.421875,
    "eval_logits/rejected": -1.390625,
    "eval_logps/chosen": -322.0,
    "eval_logps/rejected": -298.0,
    "eval_loss": 0.6962363123893738,
    "eval_nll_loss": 1.046875,
    "eval_ref_probs/chosen": 0.0038271306548267603,
    "eval_ref_probs/rejected": 0.0034569373819977045,
    "eval_rejected_term/max": 0.00031893892446532845,
    "eval_rejected_term/min": 0.00031893892446532845,
    "eval_rewards/accuracies": 0.6359999775886536,
    "eval_rewards/chosen": -0.01556396484375,
    "eval_rewards/margins": 0.1435546875,
    "eval_rewards/rejected": -0.1591796875,
    "eval_runtime": 25.5759,
    "eval_samples_per_second": 39.099,
    "eval_steps_per_second": 9.775
}