eval_results.json

{
    "epoch": 2.999098751126561,
    "eval_kl/chosen": 0.00026103382697328925,
    "eval_kl/rejected": 0.0006104370695538819,
    "eval_logits/chosen": -1.4609375,
    "eval_logits/rejected": -1.4296875,
    "eval_logps/chosen": -322.0,
    "eval_logps/rejected": -298.0,
    "eval_loss": 0.678048849105835,
    "eval_nll_loss": 1.046875,
    "eval_ref_probs/chosen": 0.0038271306548267603,
    "eval_ref_probs/rejected": 0.0034569373819977045,
    "eval_rejected_term/max": 0.0006104370695538819,
    "eval_rejected_term/min": 0.0006104370695538819,
    "eval_rewards/accuracies": 0.6600000262260437,
    "eval_rewards/chosen": -0.026611328125,
    "eval_rewards/margins": 0.1748046875,
    "eval_rewards/rejected": -0.201171875,
    "eval_runtime": 32.0895,
    "eval_samples_per_second": 31.163,
    "eval_steps_per_second": 7.791
}