eval_results.json

{
    "epoch": 2.999098751126561,
    "eval_kl/chosen": 0.000735566602088511,
    "eval_kl/rejected": -0.00039990298682823777,
    "eval_logits/chosen": -1.15625,
    "eval_logits/rejected": -1.125,
    "eval_logps/chosen": -328.0,
    "eval_logps/rejected": -300.0,
    "eval_loss": 1.1639667749404907,
    "eval_nll_loss": 1.0703125,
    "eval_ref_probs/chosen": 0.0038271306548267603,
    "eval_ref_probs/rejected": 0.0034569373819977045,
    "eval_rejected_term/max": -0.00039990298682823777,
    "eval_rejected_term/min": -0.00039990298682823777,
    "eval_rewards/accuracies": 0.3319999873638153,
    "eval_rewards/chosen": -0.73828125,
    "eval_rewards/margins": -0.34375,
    "eval_rewards/rejected": -0.396484375,
    "eval_runtime": 25.5284,
    "eval_samples_per_second": 39.172,
    "eval_steps_per_second": 9.793
}