eval_results.json

{
    "epoch": 90.0,
    "eval_kl/both": 0.0018794680945575237,
    "eval_kl/chosen": 0.0019237716915085912,
    "eval_kl/rejected": -4.440525299287401e-05,
    "eval_logits/chosen": -1.734375,
    "eval_logits/rejected": -1.71875,
    "eval_logps/chosen": -532.0,
    "eval_logps/rejected": -458.0,
    "eval_loss": 111.23272705078125,
    "eval_nll_loss": 1.8515625,
    "eval_ref_probs/chosen": 0.0038271306548267603,
    "eval_ref_probs/rejected": 0.0034569373819977045,
    "eval_rejected_term/max": -4.440525299287401e-05,
    "eval_rejected_term/min": -4.440525299287401e-05,
    "eval_rewards/accuracies": 0.36000001430511475,
    "eval_rewards/chosen": -21.125,
    "eval_rewards/margins": -4.875,
    "eval_rewards/rejected": -16.25,
    "eval_runtime": 25.7368,
    "eval_samples_per_second": 38.855,
    "eval_steps_per_second": 9.714
}