{ "step": 1000, "metrics": { "eval_rew_align/loss_libero_90": 0.07622170671820641, "eval_rew_align/pearson_libero_90": 0.9599514076421588, "eval_p_rank/kendall_last_libero_90": 0.4991780821917808, "eval_p_rank/kendall_rewind_last_libero_90": 0.8356164383561644, "eval_p_rank/avg_succ_fail_diff_last_libero_90": 0.08843660197437626, "eval_p_rank/min_succ_fail_diff_last_libero_90": -0.08279139995574952, "eval_p_rank/max_succ_fail_diff_last_libero_90": 0.2787765145301819, "eval_p_rank/ranking_acc_last_libero_90": 0.7495890410958904, "eval_p_rank/ranking_acc_all_pairs_last_libero_90": 0.7495890410958904, "eval_p_rank/ranking_acc_failure_vs_successful_last_libero_90": 0.7495890410958904, "eval_p_rank/kendall_avg_libero_90": 0.4991780821917808, "eval_p_rank/kendall_rewind_avg_libero_90": 0.8356164383561644, "eval_p_rank/avg_succ_fail_diff_avg_libero_90": 0.08843660197437626, "eval_p_rank/min_succ_fail_diff_avg_libero_90": -0.08279139995574952, "eval_p_rank/max_succ_fail_diff_avg_libero_90": 0.2787765145301819, "eval_p_rank/ranking_acc_avg_libero_90": 0.7495890410958904, "eval_p_rank/ranking_acc_all_pairs_avg_libero_90": 0.7495890410958904, "eval_p_rank/ranking_acc_failure_vs_successful_avg_libero_90": 0.7495890410958904, "eval_p_rank/kendall_sum_libero_90": 0.4991780821917808, "eval_p_rank/kendall_rewind_sum_libero_90": 0.8356164383561644, "eval_p_rank/avg_succ_fail_diff_sum_libero_90": 0.08843660197437626, "eval_p_rank/min_succ_fail_diff_sum_libero_90": -0.08279139995574952, "eval_p_rank/max_succ_fail_diff_sum_libero_90": 0.2787765145301819, "eval_p_rank/ranking_acc_sum_libero_90": 0.7495890410958904, "eval_p_rank/ranking_acc_all_pairs_sum_libero_90": 0.7495890410958904, "eval_p_rank/ranking_acc_failure_vs_successful_sum_libero_90": 0.7495890410958904, "eval_rew_align/loss_libero_10": 0.057464000582695005, "eval_rew_align/pearson_libero_10": 0.9774006704323236, "eval_p_rank/kendall_last_libero_10": 0.76, "eval_p_rank/kendall_rewind_last_libero_10": 1.0, "eval_p_rank/avg_succ_fail_diff_last_libero_10": 0.14927782282233237, "eval_p_rank/min_succ_fail_diff_last_libero_10": 0.02288996577262875, "eval_p_rank/max_succ_fail_diff_last_libero_10": 0.20963821858167647, "eval_p_rank/ranking_acc_last_libero_10": 0.88, "eval_p_rank/ranking_acc_all_pairs_last_libero_10": 0.88, "eval_p_rank/ranking_acc_failure_vs_successful_last_libero_10": 0.88, "eval_p_rank/kendall_avg_libero_10": 0.76, "eval_p_rank/kendall_rewind_avg_libero_10": 1.0, "eval_p_rank/avg_succ_fail_diff_avg_libero_10": 0.14927782282233237, "eval_p_rank/min_succ_fail_diff_avg_libero_10": 0.02288996577262875, "eval_p_rank/max_succ_fail_diff_avg_libero_10": 0.20963821858167647, "eval_p_rank/ranking_acc_avg_libero_10": 0.88, "eval_p_rank/ranking_acc_all_pairs_avg_libero_10": 0.88, "eval_p_rank/ranking_acc_failure_vs_successful_avg_libero_10": 0.88, "eval_p_rank/kendall_sum_libero_10": 0.76, "eval_p_rank/kendall_rewind_sum_libero_10": 1.0, "eval_p_rank/avg_succ_fail_diff_sum_libero_10": 0.14927782282233237, "eval_p_rank/min_succ_fail_diff_sum_libero_10": 0.02288996577262875, "eval_p_rank/max_succ_fail_diff_sum_libero_10": 0.20963821858167647, "eval_p_rank/ranking_acc_sum_libero_10": 0.88, "eval_p_rank/ranking_acc_all_pairs_sum_libero_10": 0.88, "eval_p_rank/ranking_acc_failure_vs_successful_sum_libero_10": 0.88, "time/custom_evaluations": 50.40835566003807 } }