deepseek-8b-orpo-full / eval_results.json
zfz1's picture
End of training
4310846 verified
{
"epoch": 2.0,
"eval_log_odds_chosen": 5.353264808654785,
"eval_log_odds_ratio": -0.020422911271452904,
"eval_logits/chosen": 8.594698905944824,
"eval_logits/rejected": 13.890603065490723,
"eval_logps/chosen": -0.08852054923772812,
"eval_logps/rejected": -2.195981502532959,
"eval_loss": 0.07974232733249664,
"eval_nll_loss": 0.07859200984239578,
"eval_rewards/accuracies": 1.0,
"eval_rewards/chosen": -0.008852055296301842,
"eval_rewards/margins": 0.21074610948562622,
"eval_rewards/rejected": -0.2195981740951538,
"eval_runtime": 336.9846,
"eval_samples": 5398,
"eval_samples_per_second": 16.019,
"eval_steps_per_second": 1.003
}