eval_results.json

{
    "epoch": 2.999098751126561,
    "eval_kl/both": 0.0004373829287942499,
    "eval_kl/chosen": 0.0002123450394719839,
    "eval_kl/rejected": 0.00022506651293952018,
    "eval_logits/chosen": -1.4140625,
    "eval_logits/rejected": -1.3828125,
    "eval_logps/chosen": -322.0,
    "eval_logps/rejected": -298.0,
    "eval_loss": 0.7687578201293945,
    "eval_nll_loss": 1.046875,
    "eval_ref_probs/chosen": 0.0038271306548267603,
    "eval_ref_probs/rejected": 0.0034569373819977045,
    "eval_rejected_term/max": 0.00022506651293952018,
    "eval_rejected_term/min": 0.00022506651293952018,
    "eval_rewards/accuracies": 0.6039999723434448,
    "eval_rewards/chosen": 0.005157470703125,
    "eval_rewards/margins": 0.12451171875,
    "eval_rewards/rejected": -0.119140625,
    "eval_runtime": 25.4994,
    "eval_samples_per_second": 39.217,
    "eval_steps_per_second": 9.804
}