{ "step": 450, "metrics": { "eval_rew_align/loss_libero_90": 0.13319981321692467, "eval_rew_align/pearson_libero_90": 0.8769948951244139, "eval_p_rank/kendall_last_libero_90": 0.7238356164383561, "eval_p_rank/kendall_rewind_last_libero_90": 0.9726027397260274, "eval_p_rank/avg_succ_fail_diff_last_libero_90": 0.19305384261661196, "eval_p_rank/min_succ_fail_diff_last_libero_90": -0.07096144258975984, "eval_p_rank/max_succ_fail_diff_last_libero_90": 0.5362282678484916, "eval_p_rank/ranking_acc_last_libero_90": 0.8619178082191781, "eval_p_rank/ranking_acc_all_pairs_last_libero_90": 0.8619178082191781, "eval_p_rank/ranking_acc_failure_vs_successful_last_libero_90": 0.8619178082191781, "eval_p_rank/kendall_avg_libero_90": 0.7238356164383561, "eval_p_rank/kendall_rewind_avg_libero_90": 0.9726027397260274, "eval_p_rank/avg_succ_fail_diff_avg_libero_90": 0.19305384261661196, "eval_p_rank/min_succ_fail_diff_avg_libero_90": -0.07096144258975984, "eval_p_rank/max_succ_fail_diff_avg_libero_90": 0.5362282678484916, "eval_p_rank/ranking_acc_avg_libero_90": 0.8619178082191781, "eval_p_rank/ranking_acc_all_pairs_avg_libero_90": 0.8619178082191781, "eval_p_rank/ranking_acc_failure_vs_successful_avg_libero_90": 0.8619178082191781, "eval_p_rank/kendall_sum_libero_90": 0.7238356164383561, "eval_p_rank/kendall_rewind_sum_libero_90": 0.9726027397260274, "eval_p_rank/avg_succ_fail_diff_sum_libero_90": 0.19305384261661196, "eval_p_rank/min_succ_fail_diff_sum_libero_90": -0.07096144258975984, "eval_p_rank/max_succ_fail_diff_sum_libero_90": 0.5362282678484916, "eval_p_rank/ranking_acc_sum_libero_90": 0.8619178082191781, "eval_p_rank/ranking_acc_all_pairs_sum_libero_90": 0.8619178082191781, "eval_p_rank/ranking_acc_failure_vs_successful_sum_libero_90": 0.8619178082191781, "eval_rew_align/loss_libero_10": 0.1326389029622078, "eval_rew_align/pearson_libero_10": 0.9187725430307424, "eval_p_rank/kendall_last_libero_10": 0.8960000000000001, "eval_p_rank/kendall_rewind_last_libero_10": 1.0, "eval_p_rank/avg_succ_fail_diff_last_libero_10": 0.20152622431516648, "eval_p_rank/min_succ_fail_diff_last_libero_10": 0.014177948236465454, "eval_p_rank/max_succ_fail_diff_last_libero_10": 0.29674754142761234, "eval_p_rank/ranking_acc_last_libero_10": 0.948, "eval_p_rank/ranking_acc_all_pairs_last_libero_10": 0.948, "eval_p_rank/ranking_acc_failure_vs_successful_last_libero_10": 0.948, "eval_p_rank/kendall_avg_libero_10": 0.8960000000000001, "eval_p_rank/kendall_rewind_avg_libero_10": 1.0, "eval_p_rank/avg_succ_fail_diff_avg_libero_10": 0.20152622431516648, "eval_p_rank/min_succ_fail_diff_avg_libero_10": 0.014177948236465454, "eval_p_rank/max_succ_fail_diff_avg_libero_10": 0.29674754142761234, "eval_p_rank/ranking_acc_avg_libero_10": 0.948, "eval_p_rank/ranking_acc_all_pairs_avg_libero_10": 0.948, "eval_p_rank/ranking_acc_failure_vs_successful_avg_libero_10": 0.948, "eval_p_rank/kendall_sum_libero_10": 0.8960000000000001, "eval_p_rank/kendall_rewind_sum_libero_10": 1.0, "eval_p_rank/avg_succ_fail_diff_sum_libero_10": 0.20152622431516648, "eval_p_rank/min_succ_fail_diff_sum_libero_10": 0.014177948236465454, "eval_p_rank/max_succ_fail_diff_sum_libero_10": 0.29674754142761234, "eval_p_rank/ranking_acc_sum_libero_10": 0.948, "eval_p_rank/ranking_acc_all_pairs_sum_libero_10": 0.948, "eval_p_rank/ranking_acc_failure_vs_successful_sum_libero_10": 0.948, "time/custom_evaluations": 51.121184929972515 } }