{ "checkpoint": { "adapter_path": "large_artifacts/qwen_hidden_vm_curriculum_repair/checkpoints/main_curriculum_trace_s512/adapter", "heads_path": "large_artifacts/qwen_hidden_vm_curriculum_repair/checkpoints/main_curriculum_trace_s512/heads.pt", "run": "main_curriculum_trace_s512" }, "metadata": { "args": { "arg_distance_temperature": 4.0, "arg_selection_loss_weight": 4.0, "arg_trace_loss_weight": 4.0, "arg_window": 16, "curriculum_schedule": "4:240,6:700", "device_map": "auto", "direct_answer_loss_weight": 0.0, "domain_eval_examples": 32, "domains": "arithmetic,calendar,unit,list,boolean,lookup", "eval_batch_size": 8, "eval_every": 700, "eval_examples": 192, "eval_length": 6, "eval_pairs": 128, "eval_seed": 104001, "executor_loss_weight": 0.2, "grad_clip": 1.0, "gradient_checkpointing": true, "hard_length": 8, "harder_length": 10, "head_width": 512, "init_selection_loss_weight": 1.0, "init_trace_loss_weight": 4.0, "load_in_4bit": true, "log_every": 70, "lora_alpha": 16, "lora_dropout": 0.05, "lora_r": 8, "lora_target_modules": "all-linear", "lr": 5e-05, "max_length": 768, "max_steps": 10, "model_id": "Qwen/Qwen3-4B", "op_selection_loss_weight": 1.0, "op_trace_loss_weight": 1.0, "rank_temperature": 1.0, "repair_eval_topk": 3, "repair_include_unchanged": true, "repair_max_edits": 2, "repair_selection_loss_weight": 0.0, "repair_source_examples": 512, "repair_steps": 0, "repair_target_mode": "repair_only", "repair_train_topk": 3, "role_pos_weight": 20.0, "run_name": "main_curriculum_trace_s512", "seed": 104, "selection_loss_weight": 1.0, "state_loss_weight": 0.05, "torch_dtype": "bf16", "trace_loss_weight": 1.0, "train_batch_size": 2, "train_examples": 512, "train_lora": true, "train_max_len": 6, "train_min_len": 1, "train_steps": 700, "train_template_mode": "mixed", "val_examples": 144, "value_modulus": 97, "variant": "trace", "weight_decay": 0.0 }, "cuda_available": true, "gpu_name": "NVIDIA RTX 6000 Ada Generation", "hidden_dim": 2560, "platform": "Linux-6.8.0-62-generic-x86_64-with-glibc2.39", "python": "3.12.3", "torch": "2.8.0+cu128", "train_seconds": 605.8040864467621 }, "metrics": { "domain_arithmetic": { "arg_accuracy": 0.90625, "direct_accuracy": 0.0, "domain_arithmetic_direct_accuracy": 0.0, "domain_arithmetic_executor_accuracy": 0.53125, "domain_arithmetic_n": 32.0, "domain_arithmetic_repair_executor_accuracy": 1.0, "executor_accuracy": 0.53125, "init_accuracy": 1.0, "n": 32.0, "op_accuracy": 0.9895833333333334, "program_exact": 0.53125, "repair_avg_candidates": 339.0, "repair_avg_verified": 1.0, "repair_changed_fraction": 0.46875, "repair_executor_accuracy": 1.0, "repair_found_fraction": 1.0, "repair_program_exact": 0.875, "repair_state_accuracy": 0.9375, "repair_state_prefix_fraction": 0.890625, "state_accuracy": 0.7447916666666666, "state_prefix_fraction": 0.7447916666666666 }, "domain_boolean": { "arg_accuracy": 0.9010416666666666, "direct_accuracy": 0.375, "domain_boolean_direct_accuracy": 0.375, "domain_boolean_executor_accuracy": 0.90625, "domain_boolean_n": 32.0, "domain_boolean_repair_executor_accuracy": 1.0, "executor_accuracy": 0.90625, "init_accuracy": 1.0, "n": 32.0, "op_accuracy": 1.0, "program_exact": 0.46875, "repair_avg_candidates": 339.0, "repair_avg_verified": 1.0, "repair_changed_fraction": 0.09375, "repair_executor_accuracy": 1.0, "repair_found_fraction": 1.0, "repair_program_exact": 0.46875, "repair_state_accuracy": 0.734375, "repair_state_prefix_fraction": 0.6458333333333334, "state_accuracy": 0.71875, "state_prefix_fraction": 0.6458333333333334 }, "domain_calendar": { "arg_accuracy": 0.9166666666666666, "direct_accuracy": 0.0625, "domain_calendar_direct_accuracy": 0.0625, "domain_calendar_executor_accuracy": 0.6875, "domain_calendar_n": 32.0, "domain_calendar_repair_executor_accuracy": 1.0, "executor_accuracy": 0.6875, "init_accuracy": 1.0, "n": 32.0, "op_accuracy": 1.0, "program_exact": 0.625, "repair_avg_candidates": 339.0, "repair_avg_verified": 1.0, "repair_changed_fraction": 0.3125, "repair_executor_accuracy": 1.0, "repair_found_fraction": 1.0, "repair_program_exact": 0.75, "repair_state_accuracy": 0.9375, "repair_state_prefix_fraction": 0.9010416666666666, "state_accuracy": 0.8489583333333334, "state_prefix_fraction": 0.8385416666666666 }, "domain_list": { "arg_accuracy": 0.9166666666666666, "direct_accuracy": 0.03125, "domain_list_direct_accuracy": 0.03125, "domain_list_executor_accuracy": 0.96875, "domain_list_n": 32.0, "domain_list_repair_executor_accuracy": 1.0, "executor_accuracy": 0.96875, "init_accuracy": 1.0, "n": 32.0, "op_accuracy": 1.0, "program_exact": 0.5625, "repair_avg_candidates": 339.0, "repair_avg_verified": 1.0, "repair_changed_fraction": 0.03125, "repair_executor_accuracy": 1.0, "repair_found_fraction": 1.0, "repair_program_exact": 0.59375, "repair_state_accuracy": 0.96875, "repair_state_prefix_fraction": 0.8697916666666666, "state_accuracy": 0.9635416666666666, "state_prefix_fraction": 0.8645833333333334 }, "domain_lookup": { "arg_accuracy": 0.9427083333333334, "direct_accuracy": 0.0, "domain_lookup_direct_accuracy": 0.0, "domain_lookup_executor_accuracy": 0.84375, "domain_lookup_n": 32.0, "domain_lookup_repair_executor_accuracy": 1.0, "executor_accuracy": 0.84375, "init_accuracy": 1.0, "n": 32.0, "op_accuracy": 1.0, "program_exact": 0.78125, "repair_avg_candidates": 339.0, "repair_avg_verified": 1.0, "repair_changed_fraction": 0.15625, "repair_executor_accuracy": 1.0, "repair_found_fraction": 1.0, "repair_program_exact": 0.78125, "repair_state_accuracy": 0.90625, "repair_state_prefix_fraction": 0.8645833333333334, "state_accuracy": 0.875, "state_prefix_fraction": 0.859375 }, "domain_unit": { "arg_accuracy": 0.9010416666666666, "direct_accuracy": 0.03125, "domain_unit_direct_accuracy": 0.03125, "domain_unit_executor_accuracy": 0.53125, "domain_unit_n": 32.0, "domain_unit_repair_executor_accuracy": 1.0, "executor_accuracy": 0.53125, "init_accuracy": 1.0, "n": 32.0, "op_accuracy": 0.984375, "program_exact": 0.53125, "repair_avg_candidates": 339.0, "repair_avg_verified": 1.0, "repair_changed_fraction": 0.46875, "repair_executor_accuracy": 1.0, "repair_found_fraction": 1.0, "repair_program_exact": 0.84375, "repair_state_accuracy": 0.921875, "repair_state_prefix_fraction": 0.875, "state_accuracy": 0.8072916666666666, "state_prefix_fraction": 0.8072916666666666 }, "fresh_paired_mixed": { "arg_accuracy": 0.9095052083333334, "compiler_pair_program_consistency": 0.703125, "compiler_pair_state_consistency": 0.7265625, "direct_accuracy": 0.109375, "domain_arithmetic_direct_accuracy": 0.0, "domain_arithmetic_executor_accuracy": 0.6136363636363636, "domain_arithmetic_n": 44.0, "domain_arithmetic_repair_executor_accuracy": 1.0, "domain_boolean_direct_accuracy": 0.42857142857142855, "domain_boolean_executor_accuracy": 0.9761904761904762, "domain_boolean_n": 42.0, "domain_boolean_repair_executor_accuracy": 1.0, "domain_calendar_direct_accuracy": 0.18181818181818182, "domain_calendar_executor_accuracy": 0.7045454545454546, "domain_calendar_n": 44.0, "domain_calendar_repair_executor_accuracy": 1.0, "domain_list_direct_accuracy": 0.047619047619047616, "domain_list_executor_accuracy": 0.7619047619047619, "domain_list_n": 42.0, "domain_list_repair_executor_accuracy": 1.0, "domain_lookup_direct_accuracy": 0.0, "domain_lookup_executor_accuracy": 0.7142857142857143, "domain_lookup_n": 42.0, "domain_lookup_repair_executor_accuracy": 0.9523809523809523, "domain_unit_direct_accuracy": 0.0, "domain_unit_executor_accuracy": 0.5476190476190477, "domain_unit_n": 42.0, "domain_unit_repair_executor_accuracy": 0.9761904761904762, "executor_accuracy": 0.71875, "executor_pair_answer_consistency": 0.7578125, "executor_pair_both_correct": 0.625, "init_accuracy": 1.0, "n": 256.0, "op_accuracy": 0.9973958333333334, "program_exact": 0.55859375, "repair_avg_candidates": 339.0, "repair_avg_verified": 0.98828125, "repair_changed_fraction": 0.26953125, "repair_executor_accuracy": 0.98828125, "repair_found_fraction": 0.98828125, "repair_pair_answer_consistency": 0.984375, "repair_pair_both_correct": 0.984375, "repair_pair_state_consistency": 0.875, "repair_program_exact": 0.67578125, "repair_state_accuracy": 0.900390625, "repair_state_prefix_fraction": 0.8072916666666666, "state_accuracy": 0.8131510416666666, "state_prefix_fraction": 0.7727864583333334 }, "fresh_paraphrase_mixed": { "arg_accuracy": 0.8723958333333334, "direct_accuracy": 0.11979166666666667, "domain_arithmetic_direct_accuracy": 0.0, "domain_arithmetic_executor_accuracy": 0.28125, "domain_arithmetic_n": 32.0, "domain_arithmetic_repair_executor_accuracy": 1.0, "domain_boolean_direct_accuracy": 0.46875, "domain_boolean_executor_accuracy": 0.90625, "domain_boolean_n": 32.0, "domain_boolean_repair_executor_accuracy": 1.0, "domain_calendar_direct_accuracy": 0.25, "domain_calendar_executor_accuracy": 0.40625, "domain_calendar_n": 32.0, "domain_calendar_repair_executor_accuracy": 1.0, "domain_list_direct_accuracy": 0.0, "domain_list_executor_accuracy": 0.84375, "domain_list_n": 32.0, "domain_list_repair_executor_accuracy": 1.0, "domain_lookup_direct_accuracy": 0.0, "domain_lookup_executor_accuracy": 0.78125, "domain_lookup_n": 32.0, "domain_lookup_repair_executor_accuracy": 0.96875, "domain_unit_direct_accuracy": 0.0, "domain_unit_executor_accuracy": 0.3125, "domain_unit_n": 32.0, "domain_unit_repair_executor_accuracy": 0.96875, "executor_accuracy": 0.5885416666666666, "init_accuracy": 1.0, "n": 192.0, "op_accuracy": 0.9921875, "program_exact": 0.4010416666666667, "repair_avg_candidates": 339.0, "repair_avg_verified": 0.9895833333333334, "repair_changed_fraction": 0.4010416666666667, "repair_executor_accuracy": 0.9895833333333334, "repair_found_fraction": 0.9895833333333334, "repair_program_exact": 0.578125, "repair_state_accuracy": 0.8498263888888888, "repair_state_prefix_fraction": 0.7439236111111112, "state_accuracy": 0.734375, "state_prefix_fraction": 0.6944444444444444 }, "fresh_standard_mixed": { "arg_accuracy": 0.9253472222222222, "direct_accuracy": 0.09375, "domain_arithmetic_direct_accuracy": 0.0, "domain_arithmetic_executor_accuracy": 0.75, "domain_arithmetic_n": 32.0, "domain_arithmetic_repair_executor_accuracy": 1.0, "domain_boolean_direct_accuracy": 0.5, "domain_boolean_executor_accuracy": 0.84375, "domain_boolean_n": 32.0, "domain_boolean_repair_executor_accuracy": 1.0, "domain_calendar_direct_accuracy": 0.0625, "domain_calendar_executor_accuracy": 0.625, "domain_calendar_n": 32.0, "domain_calendar_repair_executor_accuracy": 1.0, "domain_list_direct_accuracy": 0.0, "domain_list_executor_accuracy": 0.75, "domain_list_n": 32.0, "domain_list_repair_executor_accuracy": 1.0, "domain_lookup_direct_accuracy": 0.0, "domain_lookup_executor_accuracy": 0.75, "domain_lookup_n": 32.0, "domain_lookup_repair_executor_accuracy": 1.0, "domain_unit_direct_accuracy": 0.0, "domain_unit_executor_accuracy": 0.6875, "domain_unit_n": 32.0, "domain_unit_repair_executor_accuracy": 1.0, "executor_accuracy": 0.734375, "init_accuracy": 1.0, "n": 192.0, "op_accuracy": 1.0, "program_exact": 0.6197916666666666, "repair_avg_candidates": 339.0, "repair_avg_verified": 1.0, "repair_changed_fraction": 0.265625, "repair_executor_accuracy": 1.0, "repair_found_fraction": 1.0, "repair_program_exact": 0.7239583333333334, "repair_state_accuracy": 0.8828125, "repair_state_prefix_fraction": 0.8177083333333334, "state_accuracy": 0.8029513888888888, "state_prefix_fraction": 0.7760416666666666 }, "hard_paraphrase_mixed": { "arg_accuracy": 0.81640625, "direct_accuracy": 0.140625, "domain_arithmetic_direct_accuracy": 0.0, "domain_arithmetic_executor_accuracy": 0.09375, "domain_arithmetic_n": 32.0, "domain_arithmetic_repair_executor_accuracy": 0.96875, "domain_boolean_direct_accuracy": 0.59375, "domain_boolean_executor_accuracy": 0.78125, "domain_boolean_n": 32.0, "domain_boolean_repair_executor_accuracy": 1.0, "domain_calendar_direct_accuracy": 0.21875, "domain_calendar_executor_accuracy": 0.3125, "domain_calendar_n": 32.0, "domain_calendar_repair_executor_accuracy": 1.0, "domain_list_direct_accuracy": 0.0, "domain_list_executor_accuracy": 0.78125, "domain_list_n": 32.0, "domain_list_repair_executor_accuracy": 1.0, "domain_lookup_direct_accuracy": 0.0, "domain_lookup_executor_accuracy": 0.5, "domain_lookup_n": 32.0, "domain_lookup_repair_executor_accuracy": 0.96875, "domain_unit_direct_accuracy": 0.03125, "domain_unit_executor_accuracy": 0.0, "domain_unit_n": 32.0, "domain_unit_repair_executor_accuracy": 1.0, "executor_accuracy": 0.4114583333333333, "init_accuracy": 1.0, "n": 192.0, "op_accuracy": 0.9498697916666666, "program_exact": 0.203125, "repair_avg_candidates": 579.0, "repair_avg_verified": 0.9895833333333334, "repair_changed_fraction": 0.578125, "repair_executor_accuracy": 0.9895833333333334, "repair_found_fraction": 0.9895833333333334, "repair_program_exact": 0.359375, "repair_state_accuracy": 0.7141927083333334, "repair_state_prefix_fraction": 0.6184895833333334, "state_accuracy": 0.6783854166666666, "state_prefix_fraction": 0.640625 }, "hard_standard_mixed": { "arg_accuracy": 0.8938802083333334, "direct_accuracy": 0.109375, "domain_arithmetic_direct_accuracy": 0.03125, "domain_arithmetic_executor_accuracy": 0.5, "domain_arithmetic_n": 32.0, "domain_arithmetic_repair_executor_accuracy": 1.0, "domain_boolean_direct_accuracy": 0.375, "domain_boolean_executor_accuracy": 0.71875, "domain_boolean_n": 32.0, "domain_boolean_repair_executor_accuracy": 1.0, "domain_calendar_direct_accuracy": 0.21875, "domain_calendar_executor_accuracy": 0.53125, "domain_calendar_n": 32.0, "domain_calendar_repair_executor_accuracy": 1.0, "domain_list_direct_accuracy": 0.03125, "domain_list_executor_accuracy": 0.625, "domain_list_n": 32.0, "domain_list_repair_executor_accuracy": 1.0, "domain_lookup_direct_accuracy": 0.0, "domain_lookup_executor_accuracy": 0.5, "domain_lookup_n": 32.0, "domain_lookup_repair_executor_accuracy": 1.0, "domain_unit_direct_accuracy": 0.0, "domain_unit_executor_accuracy": 0.4375, "domain_unit_n": 32.0, "domain_unit_repair_executor_accuracy": 1.0, "executor_accuracy": 0.5520833333333334, "init_accuracy": 1.0, "n": 192.0, "op_accuracy": 0.9889322916666666, "program_exact": 0.3802083333333333, "repair_avg_candidates": 579.0, "repair_avg_verified": 1.0, "repair_changed_fraction": 0.4479166666666667, "repair_executor_accuracy": 1.0, "repair_found_fraction": 1.0, "repair_program_exact": 0.5625, "repair_state_accuracy": 0.8313802083333334, "repair_state_prefix_fraction": 0.7317708333333334, "state_accuracy": 0.7408854166666666, "state_prefix_fraction": 0.701171875 }, "harder_paraphrase_mixed": { "arg_accuracy": 0.7197916666666667, "direct_accuracy": 0.09375, "domain_arithmetic_direct_accuracy": 0.03125, "domain_arithmetic_executor_accuracy": 0.0, "domain_arithmetic_n": 32.0, "domain_arithmetic_repair_executor_accuracy": 1.0, "domain_boolean_direct_accuracy": 0.3125, "domain_boolean_executor_accuracy": 0.25, "domain_boolean_n": 32.0, "domain_boolean_repair_executor_accuracy": 1.0, "domain_calendar_direct_accuracy": 0.1875, "domain_calendar_executor_accuracy": 0.1875, "domain_calendar_n": 32.0, "domain_calendar_repair_executor_accuracy": 1.0, "domain_list_direct_accuracy": 0.0, "domain_list_executor_accuracy": 0.375, "domain_list_n": 32.0, "domain_list_repair_executor_accuracy": 1.0, "domain_lookup_direct_accuracy": 0.0, "domain_lookup_executor_accuracy": 0.1875, "domain_lookup_n": 32.0, "domain_lookup_repair_executor_accuracy": 0.9375, "domain_unit_direct_accuracy": 0.03125, "domain_unit_executor_accuracy": 0.0, "domain_unit_n": 32.0, "domain_unit_repair_executor_accuracy": 1.0, "executor_accuracy": 0.16666666666666666, "init_accuracy": 1.0, "n": 192.0, "op_accuracy": 0.8536458333333333, "program_exact": 0.03125, "repair_avg_candidates": 883.0, "repair_avg_verified": 0.9895833333333334, "repair_changed_fraction": 0.8229166666666666, "repair_executor_accuracy": 0.9895833333333334, "repair_found_fraction": 0.9895833333333334, "repair_program_exact": 0.08333333333333333, "repair_state_accuracy": 0.578125, "repair_state_prefix_fraction": 0.44635416666666666, "state_accuracy": 0.5640625, "state_prefix_fraction": 0.5203125 }, "harder_standard_mixed": { "arg_accuracy": 0.8270833333333333, "direct_accuracy": 0.109375, "domain_arithmetic_direct_accuracy": 0.03125, "domain_arithmetic_executor_accuracy": 0.03125, "domain_arithmetic_n": 32.0, "domain_arithmetic_repair_executor_accuracy": 1.0, "domain_boolean_direct_accuracy": 0.5, "domain_boolean_executor_accuracy": 0.375, "domain_boolean_n": 32.0, "domain_boolean_repair_executor_accuracy": 1.0, "domain_calendar_direct_accuracy": 0.09375, "domain_calendar_executor_accuracy": 0.40625, "domain_calendar_n": 32.0, "domain_calendar_repair_executor_accuracy": 1.0, "domain_list_direct_accuracy": 0.0, "domain_list_executor_accuracy": 0.5625, "domain_list_n": 32.0, "domain_list_repair_executor_accuracy": 1.0, "domain_lookup_direct_accuracy": 0.0, "domain_lookup_executor_accuracy": 0.15625, "domain_lookup_n": 32.0, "domain_lookup_repair_executor_accuracy": 0.96875, "domain_unit_direct_accuracy": 0.03125, "domain_unit_executor_accuracy": 0.28125, "domain_unit_n": 32.0, "domain_unit_repair_executor_accuracy": 1.0, "executor_accuracy": 0.3020833333333333, "init_accuracy": 1.0, "n": 192.0, "op_accuracy": 0.9416666666666667, "program_exact": 0.140625, "repair_avg_candidates": 883.0, "repair_avg_verified": 0.9947916666666666, "repair_changed_fraction": 0.6927083333333334, "repair_executor_accuracy": 0.9947916666666666, "repair_found_fraction": 0.9947916666666666, "repair_program_exact": 0.2916666666666667, "repair_state_accuracy": 0.7114583333333333, "repair_state_prefix_fraction": 0.6, "state_accuracy": 0.646875, "state_prefix_fraction": 0.6098958333333333 }, "val_mixed": { "arg_accuracy": 0.9178240740740741, "direct_accuracy": 0.10416666666666667, "domain_arithmetic_direct_accuracy": 0.0, "domain_arithmetic_executor_accuracy": 0.4583333333333333, "domain_arithmetic_n": 24.0, "domain_arithmetic_repair_executor_accuracy": 1.0, "domain_boolean_direct_accuracy": 0.375, "domain_boolean_executor_accuracy": 0.8333333333333334, "domain_boolean_n": 24.0, "domain_boolean_repair_executor_accuracy": 1.0, "domain_calendar_direct_accuracy": 0.25, "domain_calendar_executor_accuracy": 0.5416666666666666, "domain_calendar_n": 24.0, "domain_calendar_repair_executor_accuracy": 1.0, "domain_list_direct_accuracy": 0.0, "domain_list_executor_accuracy": 0.875, "domain_list_n": 24.0, "domain_list_repair_executor_accuracy": 1.0, "domain_lookup_direct_accuracy": 0.0, "domain_lookup_executor_accuracy": 0.9166666666666666, "domain_lookup_n": 24.0, "domain_lookup_repair_executor_accuracy": 1.0, "domain_unit_direct_accuracy": 0.0, "domain_unit_executor_accuracy": 0.5416666666666666, "domain_unit_n": 24.0, "domain_unit_repair_executor_accuracy": 1.0, "executor_accuracy": 0.6944444444444444, "init_accuracy": 1.0, "n": 144.0, "op_accuracy": 0.9976851851851852, "program_exact": 0.5833333333333334, "repair_avg_candidates": 339.0, "repair_avg_verified": 1.0, "repair_changed_fraction": 0.3055555555555556, "repair_executor_accuracy": 1.0, "repair_found_fraction": 1.0, "repair_program_exact": 0.7291666666666666, "repair_state_accuracy": 0.9108796296296297, "repair_state_prefix_fraction": 0.8495370370370371, "state_accuracy": 0.8067129629629629, "state_prefix_fraction": 0.7800925925925926 } }, "train_log": [ { "domain_arithmetic_arg_accuracy": 0.4375, "domain_arithmetic_direct_accuracy": 0.0, "domain_arithmetic_domain_arithmetic_direct_accuracy": 0.0, "domain_arithmetic_domain_arithmetic_executor_accuracy": 0.0, "domain_arithmetic_domain_arithmetic_n": 32.0, "domain_arithmetic_domain_arithmetic_repair_executor_accuracy": 0.90625, "domain_arithmetic_executor_accuracy": 0.0, "domain_arithmetic_init_accuracy": 0.21875, "domain_arithmetic_op_accuracy": 0.7239583333333334, "domain_arithmetic_program_exact": 0.0, "domain_arithmetic_repair_avg_candidates": 339.0, "domain_arithmetic_repair_avg_verified": 0.90625, "domain_arithmetic_repair_changed_fraction": 0.90625, "domain_arithmetic_repair_executor_accuracy": 0.90625, "domain_arithmetic_repair_found_fraction": 0.90625, "domain_arithmetic_repair_program_exact": 0.0, "domain_arithmetic_repair_state_accuracy": 0.3125, "domain_arithmetic_repair_state_prefix_fraction": 0.010416666666666666, "domain_arithmetic_state_accuracy": 0.020833333333333332, "domain_arithmetic_state_prefix_fraction": 0.020833333333333332, "domain_boolean_arg_accuracy": 0.4375, "domain_boolean_direct_accuracy": 0.59375, "domain_boolean_domain_boolean_direct_accuracy": 0.59375, "domain_boolean_domain_boolean_executor_accuracy": 0.46875, "domain_boolean_domain_boolean_n": 32.0, "domain_boolean_domain_boolean_repair_executor_accuracy": 1.0, "domain_boolean_executor_accuracy": 0.46875, "domain_boolean_init_accuracy": 0.0625, "domain_boolean_op_accuracy": 0.6302083333333334, "domain_boolean_program_exact": 0.0, "domain_boolean_repair_avg_candidates": 339.0, "domain_boolean_repair_avg_verified": 1.0, "domain_boolean_repair_changed_fraction": 0.53125, "domain_boolean_repair_executor_accuracy": 1.0, "domain_boolean_repair_found_fraction": 1.0, "domain_boolean_repair_program_exact": 0.0, "domain_boolean_repair_state_accuracy": 0.234375, "domain_boolean_repair_state_prefix_fraction": 0.036458333333333336, "domain_boolean_state_accuracy": 0.15104166666666666, "domain_boolean_state_prefix_fraction": 0.036458333333333336, "domain_calendar_arg_accuracy": 0.3645833333333333, "domain_calendar_direct_accuracy": 0.0625, "domain_calendar_domain_calendar_direct_accuracy": 0.0625, "domain_calendar_domain_calendar_executor_accuracy": 0.15625, "domain_calendar_domain_calendar_n": 32.0, "domain_calendar_domain_calendar_repair_executor_accuracy": 1.0, "domain_calendar_executor_accuracy": 0.15625, "domain_calendar_init_accuracy": 0.15625, "domain_calendar_op_accuracy": 0.7916666666666666, "domain_calendar_program_exact": 0.0, "domain_calendar_repair_avg_candidates": 339.0, "domain_calendar_repair_avg_verified": 1.0, "domain_calendar_repair_changed_fraction": 0.84375, "domain_calendar_repair_executor_accuracy": 1.0, "domain_calendar_repair_found_fraction": 1.0, "domain_calendar_repair_program_exact": 0.0, "domain_calendar_repair_state_accuracy": 0.484375, "domain_calendar_repair_state_prefix_fraction": 0.06770833333333333, "domain_calendar_state_accuracy": 0.16666666666666666, "domain_calendar_state_prefix_fraction": 0.07291666666666667, "domain_list_arg_accuracy": 0.609375, "domain_list_direct_accuracy": 0.0, "domain_list_domain_list_direct_accuracy": 0.0, "domain_list_domain_list_executor_accuracy": 0.6875, "domain_list_domain_list_n": 32.0, "domain_list_domain_list_repair_executor_accuracy": 1.0, "domain_list_executor_accuracy": 0.6875, "domain_list_init_accuracy": 0.03125, "domain_list_op_accuracy": 0.828125, "domain_list_program_exact": 0.0, "domain_list_repair_avg_candidates": 339.0, "domain_list_repair_avg_verified": 1.0, "domain_list_repair_changed_fraction": 0.3125, "domain_list_repair_executor_accuracy": 1.0, "domain_list_repair_found_fraction": 1.0, "domain_list_repair_program_exact": 0.0, "domain_list_repair_state_accuracy": 0.609375, "domain_list_repair_state_prefix_fraction": 0.22395833333333334, "domain_list_state_accuracy": 0.4791666666666667, "domain_list_state_prefix_fraction": 0.19270833333333334, "domain_lookup_arg_accuracy": 0.046875, "domain_lookup_direct_accuracy": 0.03125, "domain_lookup_domain_lookup_direct_accuracy": 0.03125, "domain_lookup_domain_lookup_executor_accuracy": 0.03125, "domain_lookup_domain_lookup_n": 32.0, "domain_lookup_domain_lookup_repair_executor_accuracy": 0.46875, "domain_lookup_executor_accuracy": 0.03125, "domain_lookup_init_accuracy": 0.0, "domain_lookup_op_accuracy": 0.265625, "domain_lookup_program_exact": 0.0, "domain_lookup_repair_avg_candidates": 339.0, "domain_lookup_repair_avg_verified": 0.46875, "domain_lookup_repair_changed_fraction": 0.4375, "domain_lookup_repair_executor_accuracy": 0.46875, "domain_lookup_repair_found_fraction": 0.46875, "domain_lookup_repair_program_exact": 0.0, "domain_lookup_repair_state_accuracy": 0.11458333333333333, "domain_lookup_repair_state_prefix_fraction": 0.026041666666666668, "domain_lookup_state_accuracy": 0.036458333333333336, "domain_lookup_state_prefix_fraction": 0.026041666666666668, "domain_unit_arg_accuracy": 0.5729166666666666, "domain_unit_direct_accuracy": 0.0, "domain_unit_domain_unit_direct_accuracy": 0.0, "domain_unit_domain_unit_executor_accuracy": 0.0, "domain_unit_domain_unit_n": 32.0, "domain_unit_domain_unit_repair_executor_accuracy": 0.9375, "domain_unit_executor_accuracy": 0.0, "domain_unit_init_accuracy": 0.21875, "domain_unit_op_accuracy": 0.8125, "domain_unit_program_exact": 0.0, "domain_unit_repair_avg_candidates": 339.0, "domain_unit_repair_avg_verified": 0.9375, "domain_unit_repair_changed_fraction": 0.9375, "domain_unit_repair_executor_accuracy": 0.9375, "domain_unit_repair_found_fraction": 0.9375, "domain_unit_repair_program_exact": 0.03125, "domain_unit_repair_state_accuracy": 0.4635416666666667, "domain_unit_repair_state_prefix_fraction": 0.18229166666666666, "domain_unit_state_accuracy": 0.08854166666666667, "domain_unit_state_prefix_fraction": 0.08854166666666667, "fresh_paired_mixed_arg_accuracy": 0.3912760416666667, "fresh_paired_mixed_compiler_pair_program_consistency": 0.1328125, "fresh_paired_mixed_compiler_pair_state_consistency": 0.1875, "fresh_paired_mixed_direct_accuracy": 0.13671875, "fresh_paired_mixed_domain_arithmetic_direct_accuracy": 0.0, "fresh_paired_mixed_domain_arithmetic_executor_accuracy": 0.022727272727272728, "fresh_paired_mixed_domain_arithmetic_n": 44.0, "fresh_paired_mixed_domain_arithmetic_repair_executor_accuracy": 0.9772727272727273, "fresh_paired_mixed_domain_boolean_direct_accuracy": 0.6190476190476191, "fresh_paired_mixed_domain_boolean_executor_accuracy": 0.4523809523809524, "fresh_paired_mixed_domain_boolean_n": 42.0, "fresh_paired_mixed_domain_boolean_repair_executor_accuracy": 1.0, "fresh_paired_mixed_domain_calendar_direct_accuracy": 0.20454545454545456, "fresh_paired_mixed_domain_calendar_executor_accuracy": 0.13636363636363635, "fresh_paired_mixed_domain_calendar_n": 44.0, "fresh_paired_mixed_domain_calendar_repair_executor_accuracy": 1.0, "fresh_paired_mixed_domain_list_direct_accuracy": 0.0, "fresh_paired_mixed_domain_list_executor_accuracy": 0.40476190476190477, "fresh_paired_mixed_domain_list_n": 42.0, "fresh_paired_mixed_domain_list_repair_executor_accuracy": 1.0, "fresh_paired_mixed_domain_lookup_direct_accuracy": 0.0, "fresh_paired_mixed_domain_lookup_executor_accuracy": 0.09523809523809523, "fresh_paired_mixed_domain_lookup_n": 42.0, "fresh_paired_mixed_domain_lookup_repair_executor_accuracy": 0.4523809523809524, "fresh_paired_mixed_domain_unit_direct_accuracy": 0.0, "fresh_paired_mixed_domain_unit_executor_accuracy": 0.0, "fresh_paired_mixed_domain_unit_n": 42.0, "fresh_paired_mixed_domain_unit_repair_executor_accuracy": 0.9047619047619048, "fresh_paired_mixed_executor_accuracy": 0.18359375, "fresh_paired_mixed_executor_pair_answer_consistency": 0.484375, "fresh_paired_mixed_executor_pair_both_correct": 0.15625, "fresh_paired_mixed_init_accuracy": 0.109375, "fresh_paired_mixed_op_accuracy": 0.6712239583333334, "fresh_paired_mixed_program_exact": 0.0, "fresh_paired_mixed_repair_avg_candidates": 339.0, "fresh_paired_mixed_repair_avg_verified": 0.890625, "fresh_paired_mixed_repair_changed_fraction": 0.70703125, "fresh_paired_mixed_repair_executor_accuracy": 0.890625, "fresh_paired_mixed_repair_found_fraction": 0.890625, "fresh_paired_mixed_repair_pair_answer_consistency": 0.90625, "fresh_paired_mixed_repair_pair_both_correct": 0.8671875, "fresh_paired_mixed_repair_pair_state_consistency": 0.2109375, "fresh_paired_mixed_repair_program_exact": 0.0, "fresh_paired_mixed_repair_state_accuracy": 0.3359375, "fresh_paired_mixed_repair_state_prefix_fraction": 0.034505208333333336, "fresh_paired_mixed_state_accuracy": 0.11653645833333333, "fresh_paired_mixed_state_prefix_fraction": 0.037760416666666664, "fresh_paraphrase_mixed_arg_accuracy": 0.4097222222222222, "fresh_paraphrase_mixed_direct_accuracy": 0.09895833333333333, "fresh_paraphrase_mixed_domain_arithmetic_direct_accuracy": 0.0, "fresh_paraphrase_mixed_domain_arithmetic_executor_accuracy": 0.0, "fresh_paraphrase_mixed_domain_arithmetic_n": 32.0, "fresh_paraphrase_mixed_domain_arithmetic_repair_executor_accuracy": 0.96875, "fresh_paraphrase_mixed_domain_boolean_direct_accuracy": 0.40625, "fresh_paraphrase_mixed_domain_boolean_executor_accuracy": 0.46875, "fresh_paraphrase_mixed_domain_boolean_n": 32.0, "fresh_paraphrase_mixed_domain_boolean_repair_executor_accuracy": 1.0, "fresh_paraphrase_mixed_domain_calendar_direct_accuracy": 0.1875, "fresh_paraphrase_mixed_domain_calendar_executor_accuracy": 0.125, "fresh_paraphrase_mixed_domain_calendar_n": 32.0, "fresh_paraphrase_mixed_domain_calendar_repair_executor_accuracy": 1.0, "fresh_paraphrase_mixed_domain_list_direct_accuracy": 0.0, "fresh_paraphrase_mixed_domain_list_executor_accuracy": 0.59375, "fresh_paraphrase_mixed_domain_list_n": 32.0, "fresh_paraphrase_mixed_domain_list_repair_executor_accuracy": 1.0, "fresh_paraphrase_mixed_domain_lookup_direct_accuracy": 0.0, "fresh_paraphrase_mixed_domain_lookup_executor_accuracy": 0.0625, "fresh_paraphrase_mixed_domain_lookup_n": 32.0, "fresh_paraphrase_mixed_domain_lookup_repair_executor_accuracy": 0.40625, "fresh_paraphrase_mixed_domain_unit_direct_accuracy": 0.0, "fresh_paraphrase_mixed_domain_unit_executor_accuracy": 0.0, "fresh_paraphrase_mixed_domain_unit_n": 32.0, "fresh_paraphrase_mixed_domain_unit_repair_executor_accuracy": 0.9375, "fresh_paraphrase_mixed_executor_accuracy": 0.20833333333333334, "fresh_paraphrase_mixed_init_accuracy": 0.125, "fresh_paraphrase_mixed_op_accuracy": 0.6744791666666666, "fresh_paraphrase_mixed_program_exact": 0.0, "fresh_paraphrase_mixed_repair_avg_candidates": 339.0, "fresh_paraphrase_mixed_repair_avg_verified": 0.8854166666666666, "fresh_paraphrase_mixed_repair_changed_fraction": 0.6770833333333334, "fresh_paraphrase_mixed_repair_executor_accuracy": 0.8854166666666666, "fresh_paraphrase_mixed_repair_found_fraction": 0.8854166666666666, "fresh_paraphrase_mixed_repair_program_exact": 0.0, "fresh_paraphrase_mixed_repair_state_accuracy": 0.3637152777777778, "fresh_paraphrase_mixed_repair_state_prefix_fraction": 0.07725694444444445, "fresh_paraphrase_mixed_state_accuracy": 0.1545138888888889, "fresh_paraphrase_mixed_state_prefix_fraction": 0.07118055555555555, "fresh_standard_mixed_arg_accuracy": 0.3611111111111111, "fresh_standard_mixed_direct_accuracy": 0.11979166666666667, "fresh_standard_mixed_domain_arithmetic_direct_accuracy": 0.0, "fresh_standard_mixed_domain_arithmetic_executor_accuracy": 0.03125, "fresh_standard_mixed_domain_arithmetic_n": 32.0, "fresh_standard_mixed_domain_arithmetic_repair_executor_accuracy": 0.875, "fresh_standard_mixed_domain_boolean_direct_accuracy": 0.59375, "fresh_standard_mixed_domain_boolean_executor_accuracy": 0.5, "fresh_standard_mixed_domain_boolean_n": 32.0, "fresh_standard_mixed_domain_boolean_repair_executor_accuracy": 1.0, "fresh_standard_mixed_domain_calendar_direct_accuracy": 0.125, "fresh_standard_mixed_domain_calendar_executor_accuracy": 0.09375, "fresh_standard_mixed_domain_calendar_n": 32.0, "fresh_standard_mixed_domain_calendar_repair_executor_accuracy": 1.0, "fresh_standard_mixed_domain_list_direct_accuracy": 0.0, "fresh_standard_mixed_domain_list_executor_accuracy": 0.625, "fresh_standard_mixed_domain_list_n": 32.0, "fresh_standard_mixed_domain_list_repair_executor_accuracy": 1.0, "fresh_standard_mixed_domain_lookup_direct_accuracy": 0.0, "fresh_standard_mixed_domain_lookup_executor_accuracy": 0.0, "fresh_standard_mixed_domain_lookup_n": 32.0, "fresh_standard_mixed_domain_lookup_repair_executor_accuracy": 0.59375, "fresh_standard_mixed_domain_unit_direct_accuracy": 0.0, "fresh_standard_mixed_domain_unit_executor_accuracy": 0.03125, "fresh_standard_mixed_domain_unit_n": 32.0, "fresh_standard_mixed_domain_unit_repair_executor_accuracy": 0.96875, "fresh_standard_mixed_executor_accuracy": 0.21354166666666666, "fresh_standard_mixed_init_accuracy": 0.036458333333333336, "fresh_standard_mixed_op_accuracy": 0.6579861111111112, "fresh_standard_mixed_program_exact": 0.0, "fresh_standard_mixed_repair_avg_candidates": 339.0, "fresh_standard_mixed_repair_avg_verified": 0.90625, "fresh_standard_mixed_repair_changed_fraction": 0.6927083333333334, "fresh_standard_mixed_repair_executor_accuracy": 0.90625, "fresh_standard_mixed_repair_found_fraction": 0.90625, "fresh_standard_mixed_repair_program_exact": 0.0, "fresh_standard_mixed_repair_state_accuracy": 0.3420138888888889, "fresh_standard_mixed_repair_state_prefix_fraction": 0.043402777777777776, "fresh_standard_mixed_state_accuracy": 0.13194444444444445, "fresh_standard_mixed_state_prefix_fraction": 0.044270833333333336, "hard_paraphrase_mixed_arg_accuracy": 0.33203125, "hard_paraphrase_mixed_direct_accuracy": 0.109375, "hard_paraphrase_mixed_domain_arithmetic_direct_accuracy": 0.0, "hard_paraphrase_mixed_domain_arithmetic_executor_accuracy": 0.0, "hard_paraphrase_mixed_domain_arithmetic_n": 32.0, "hard_paraphrase_mixed_domain_arithmetic_repair_executor_accuracy": 1.0, "hard_paraphrase_mixed_domain_boolean_direct_accuracy": 0.4375, "hard_paraphrase_mixed_domain_boolean_executor_accuracy": 0.59375, "hard_paraphrase_mixed_domain_boolean_n": 32.0, "hard_paraphrase_mixed_domain_boolean_repair_executor_accuracy": 1.0, "hard_paraphrase_mixed_domain_calendar_direct_accuracy": 0.21875, "hard_paraphrase_mixed_domain_calendar_executor_accuracy": 0.1875, "hard_paraphrase_mixed_domain_calendar_n": 32.0, "hard_paraphrase_mixed_domain_calendar_repair_executor_accuracy": 1.0, "hard_paraphrase_mixed_domain_list_direct_accuracy": 0.0, "hard_paraphrase_mixed_domain_list_executor_accuracy": 0.65625, "hard_paraphrase_mixed_domain_list_n": 32.0, "hard_paraphrase_mixed_domain_list_repair_executor_accuracy": 1.0, "hard_paraphrase_mixed_domain_lookup_direct_accuracy": 0.0, "hard_paraphrase_mixed_domain_lookup_executor_accuracy": 0.0, "hard_paraphrase_mixed_domain_lookup_n": 32.0, "hard_paraphrase_mixed_domain_lookup_repair_executor_accuracy": 0.6875, "hard_paraphrase_mixed_domain_unit_direct_accuracy": 0.0, "hard_paraphrase_mixed_domain_unit_executor_accuracy": 0.0, "hard_paraphrase_mixed_domain_unit_n": 32.0, "hard_paraphrase_mixed_domain_unit_repair_executor_accuracy": 1.0, "hard_paraphrase_mixed_executor_accuracy": 0.23958333333333334, "hard_paraphrase_mixed_init_accuracy": 0.08333333333333333, "hard_paraphrase_mixed_op_accuracy": 0.646484375, "hard_paraphrase_mixed_program_exact": 0.0, "hard_paraphrase_mixed_repair_avg_candidates": 579.0, "hard_paraphrase_mixed_repair_avg_verified": 0.9479166666666666, "hard_paraphrase_mixed_repair_changed_fraction": 0.7083333333333334, "hard_paraphrase_mixed_repair_executor_accuracy": 0.9479166666666666, "hard_paraphrase_mixed_repair_found_fraction": 0.9479166666666666, "hard_paraphrase_mixed_repair_program_exact": 0.0, "hard_paraphrase_mixed_repair_state_accuracy": 0.2994791666666667, "hard_paraphrase_mixed_repair_state_prefix_fraction": 0.044921875, "hard_paraphrase_mixed_state_accuracy": 0.14908854166666666, "hard_paraphrase_mixed_state_prefix_fraction": 0.051432291666666664, "hard_standard_mixed_arg_accuracy": 0.3430989583333333, "hard_standard_mixed_direct_accuracy": 0.140625, "hard_standard_mixed_domain_arithmetic_direct_accuracy": 0.0, "hard_standard_mixed_domain_arithmetic_executor_accuracy": 0.0, "hard_standard_mixed_domain_arithmetic_n": 32.0, "hard_standard_mixed_domain_arithmetic_repair_executor_accuracy": 1.0, "hard_standard_mixed_domain_boolean_direct_accuracy": 0.6875, "hard_standard_mixed_domain_boolean_executor_accuracy": 0.375, "hard_standard_mixed_domain_boolean_n": 32.0, "hard_standard_mixed_domain_boolean_repair_executor_accuracy": 1.0, "hard_standard_mixed_domain_calendar_direct_accuracy": 0.125, "hard_standard_mixed_domain_calendar_executor_accuracy": 0.21875, "hard_standard_mixed_domain_calendar_n": 32.0, "hard_standard_mixed_domain_calendar_repair_executor_accuracy": 1.0, "hard_standard_mixed_domain_list_direct_accuracy": 0.03125, "hard_standard_mixed_domain_list_executor_accuracy": 0.75, "hard_standard_mixed_domain_list_n": 32.0, "hard_standard_mixed_domain_list_repair_executor_accuracy": 1.0, "hard_standard_mixed_domain_lookup_direct_accuracy": 0.0, "hard_standard_mixed_domain_lookup_executor_accuracy": 0.03125, "hard_standard_mixed_domain_lookup_n": 32.0, "hard_standard_mixed_domain_lookup_repair_executor_accuracy": 0.65625, "hard_standard_mixed_domain_unit_direct_accuracy": 0.0, "hard_standard_mixed_domain_unit_executor_accuracy": 0.0, "hard_standard_mixed_domain_unit_n": 32.0, "hard_standard_mixed_domain_unit_repair_executor_accuracy": 1.0, "hard_standard_mixed_executor_accuracy": 0.22916666666666666, "hard_standard_mixed_init_accuracy": 0.046875, "hard_standard_mixed_op_accuracy": 0.6380208333333334, "hard_standard_mixed_program_exact": 0.0, "hard_standard_mixed_repair_avg_candidates": 579.0, "hard_standard_mixed_repair_avg_verified": 0.9427083333333334, "hard_standard_mixed_repair_changed_fraction": 0.7135416666666666, "hard_standard_mixed_repair_executor_accuracy": 0.9427083333333334, "hard_standard_mixed_repair_found_fraction": 0.9427083333333334, "hard_standard_mixed_repair_program_exact": 0.0, "hard_standard_mixed_repair_state_accuracy": 0.3040364583333333, "hard_standard_mixed_repair_state_prefix_fraction": 0.042317708333333336, "hard_standard_mixed_state_accuracy": 0.14322916666666666, "hard_standard_mixed_state_prefix_fraction": 0.050130208333333336, "harder_paraphrase_mixed_arg_accuracy": 0.2947916666666667, "harder_paraphrase_mixed_direct_accuracy": 0.10416666666666667, "harder_paraphrase_mixed_domain_arithmetic_direct_accuracy": 0.0, "harder_paraphrase_mixed_domain_arithmetic_executor_accuracy": 0.0, "harder_paraphrase_mixed_domain_arithmetic_n": 32.0, "harder_paraphrase_mixed_domain_arithmetic_repair_executor_accuracy": 0.96875, "harder_paraphrase_mixed_domain_boolean_direct_accuracy": 0.53125, "harder_paraphrase_mixed_domain_boolean_executor_accuracy": 0.3125, "harder_paraphrase_mixed_domain_boolean_n": 32.0, "harder_paraphrase_mixed_domain_boolean_repair_executor_accuracy": 1.0, "harder_paraphrase_mixed_domain_calendar_direct_accuracy": 0.0625, "harder_paraphrase_mixed_domain_calendar_executor_accuracy": 0.125, "harder_paraphrase_mixed_domain_calendar_n": 32.0, "harder_paraphrase_mixed_domain_calendar_repair_executor_accuracy": 1.0, "harder_paraphrase_mixed_domain_list_direct_accuracy": 0.0, "harder_paraphrase_mixed_domain_list_executor_accuracy": 0.6875, "harder_paraphrase_mixed_domain_list_n": 32.0, "harder_paraphrase_mixed_domain_list_repair_executor_accuracy": 1.0, "harder_paraphrase_mixed_domain_lookup_direct_accuracy": 0.0, "harder_paraphrase_mixed_domain_lookup_executor_accuracy": 0.03125, "harder_paraphrase_mixed_domain_lookup_n": 32.0, "harder_paraphrase_mixed_domain_lookup_repair_executor_accuracy": 0.875, "harder_paraphrase_mixed_domain_unit_direct_accuracy": 0.03125, "harder_paraphrase_mixed_domain_unit_executor_accuracy": 0.0, "harder_paraphrase_mixed_domain_unit_n": 32.0, "harder_paraphrase_mixed_domain_unit_repair_executor_accuracy": 1.0, "harder_paraphrase_mixed_executor_accuracy": 0.19270833333333334, "harder_paraphrase_mixed_init_accuracy": 0.07291666666666667, "harder_paraphrase_mixed_op_accuracy": 0.5973958333333333, "harder_paraphrase_mixed_program_exact": 0.0, "harder_paraphrase_mixed_repair_avg_candidates": 883.0, "harder_paraphrase_mixed_repair_avg_verified": 0.9739583333333334, "harder_paraphrase_mixed_repair_changed_fraction": 0.78125, "harder_paraphrase_mixed_repair_executor_accuracy": 0.9739583333333334, "harder_paraphrase_mixed_repair_found_fraction": 0.9739583333333334, "harder_paraphrase_mixed_repair_program_exact": 0.0, "harder_paraphrase_mixed_repair_state_accuracy": 0.27708333333333335, "harder_paraphrase_mixed_repair_state_prefix_fraction": 0.03854166666666667, "harder_paraphrase_mixed_state_accuracy": 0.13229166666666667, "harder_paraphrase_mixed_state_prefix_fraction": 0.0421875, "harder_standard_mixed_arg_accuracy": 0.2947916666666667, "harder_standard_mixed_direct_accuracy": 0.09895833333333333, "harder_standard_mixed_domain_arithmetic_direct_accuracy": 0.0, "harder_standard_mixed_domain_arithmetic_executor_accuracy": 0.0, "harder_standard_mixed_domain_arithmetic_n": 32.0, "harder_standard_mixed_domain_arithmetic_repair_executor_accuracy": 1.0, "harder_standard_mixed_domain_boolean_direct_accuracy": 0.4375, "harder_standard_mixed_domain_boolean_executor_accuracy": 0.5, "harder_standard_mixed_domain_boolean_n": 32.0, "harder_standard_mixed_domain_boolean_repair_executor_accuracy": 1.0, "harder_standard_mixed_domain_calendar_direct_accuracy": 0.15625, "harder_standard_mixed_domain_calendar_executor_accuracy": 0.09375, "harder_standard_mixed_domain_calendar_n": 32.0, "harder_standard_mixed_domain_calendar_repair_executor_accuracy": 1.0, "harder_standard_mixed_domain_list_direct_accuracy": 0.0, "harder_standard_mixed_domain_list_executor_accuracy": 0.71875, "harder_standard_mixed_domain_list_n": 32.0, "harder_standard_mixed_domain_list_repair_executor_accuracy": 1.0, "harder_standard_mixed_domain_lookup_direct_accuracy": 0.0, "harder_standard_mixed_domain_lookup_executor_accuracy": 0.125, "harder_standard_mixed_domain_lookup_n": 32.0, "harder_standard_mixed_domain_lookup_repair_executor_accuracy": 0.9375, "harder_standard_mixed_domain_unit_direct_accuracy": 0.0, "harder_standard_mixed_domain_unit_executor_accuracy": 0.0, "harder_standard_mixed_domain_unit_n": 32.0, "harder_standard_mixed_domain_unit_repair_executor_accuracy": 1.0, "harder_standard_mixed_executor_accuracy": 0.23958333333333334, "harder_standard_mixed_init_accuracy": 0.057291666666666664, "harder_standard_mixed_op_accuracy": 0.6166666666666667, "harder_standard_mixed_program_exact": 0.0, "harder_standard_mixed_repair_avg_candidates": 883.0, "harder_standard_mixed_repair_avg_verified": 0.9895833333333334, "harder_standard_mixed_repair_changed_fraction": 0.75, "harder_standard_mixed_repair_executor_accuracy": 0.9895833333333334, "harder_standard_mixed_repair_found_fraction": 0.9895833333333334, "harder_standard_mixed_repair_program_exact": 0.0, "harder_standard_mixed_repair_state_accuracy": 0.265625, "harder_standard_mixed_repair_state_prefix_fraction": 0.024479166666666666, "harder_standard_mixed_state_accuracy": 0.1203125, "harder_standard_mixed_state_prefix_fraction": 0.0265625, "phase": "baseline", "step": 0, "val_mixed_arg_accuracy": 0.38425925925925924, "val_mixed_direct_accuracy": 0.0763888888888889, "val_mixed_domain_arithmetic_direct_accuracy": 0.0, "val_mixed_domain_arithmetic_executor_accuracy": 0.0, "val_mixed_domain_arithmetic_n": 24.0, "val_mixed_domain_arithmetic_repair_executor_accuracy": 0.9583333333333334, "val_mixed_domain_boolean_direct_accuracy": 0.375, "val_mixed_domain_boolean_executor_accuracy": 0.5416666666666666, "val_mixed_domain_boolean_n": 24.0, "val_mixed_domain_boolean_repair_executor_accuracy": 1.0, "val_mixed_domain_calendar_direct_accuracy": 0.08333333333333333, "val_mixed_domain_calendar_executor_accuracy": 0.08333333333333333, "val_mixed_domain_calendar_n": 24.0, "val_mixed_domain_calendar_repair_executor_accuracy": 1.0, "val_mixed_domain_list_direct_accuracy": 0.0, "val_mixed_domain_list_executor_accuracy": 0.7083333333333334, "val_mixed_domain_list_n": 24.0, "val_mixed_domain_list_repair_executor_accuracy": 1.0, "val_mixed_domain_lookup_direct_accuracy": 0.0, "val_mixed_domain_lookup_executor_accuracy": 0.0, "val_mixed_domain_lookup_n": 24.0, "val_mixed_domain_lookup_repair_executor_accuracy": 0.3333333333333333, "val_mixed_domain_unit_direct_accuracy": 0.0, "val_mixed_domain_unit_executor_accuracy": 0.0, "val_mixed_domain_unit_n": 24.0, "val_mixed_domain_unit_repair_executor_accuracy": 0.8333333333333334, "val_mixed_executor_accuracy": 0.2222222222222222, "val_mixed_init_accuracy": 0.10416666666666667, "val_mixed_op_accuracy": 0.6909722222222222, "val_mixed_program_exact": 0.0, "val_mixed_repair_avg_candidates": 339.0, "val_mixed_repair_avg_verified": 0.8541666666666666, "val_mixed_repair_changed_fraction": 0.6319444444444444, "val_mixed_repair_executor_accuracy": 0.8541666666666666, "val_mixed_repair_found_fraction": 0.8541666666666666, "val_mixed_repair_program_exact": 0.0, "val_mixed_repair_state_accuracy": 0.3402777777777778, "val_mixed_repair_state_prefix_fraction": 0.041666666666666664, "val_mixed_state_accuracy": 0.14583333333333334, "val_mixed_state_prefix_fraction": 0.04050925925925926 }, { "arg_attn_loss": 0.8574167490005493, "arg_loss": 0.587339460849762, "executor_loss": 2.1179358959198, "init_attn_loss": 0.10347934067249298, "init_loss": 0.008702053688466549, "loss": 9.277847290039062, "op_attn_loss": 1.929778814315796, "op_loss": 0.4255327880382538, "phase": "curriculum_train", "role_loss": 0.509459912776947, "stage_max_len": 4, "state_loss": 1.4435406923294067, "state_train_accuracy": 0.7142857313156128, "step": 70 }, { "arg_attn_loss": 0.3759680986404419, "arg_loss": 0.2732236385345459, "executor_loss": 0.4447067081928253, "init_attn_loss": 0.0022264947183430195, "init_loss": 0.000994834816083312, "loss": 3.9375107288360596, "op_attn_loss": 0.8273237943649292, "op_loss": 0.2846919298171997, "phase": "curriculum_train", "role_loss": 0.11208517849445343, "stage_max_len": 4, "state_loss": 0.4299156069755554, "state_train_accuracy": 1.0, "step": 140 }, { "arg_attn_loss": 0.598007082939148, "arg_loss": 0.5865567922592163, "executor_loss": 1.0267366170883179, "init_attn_loss": 0.0027838617097586393, "init_loss": 1.1384418030502275e-05, "loss": 6.493476390838623, "op_attn_loss": 0.9280757904052734, "op_loss": 0.4459085464477539, "phase": "curriculum_train", "role_loss": 0.11583157628774643, "stage_max_len": 4, "state_loss": 1.1445621252059937, "state_train_accuracy": 0.800000011920929, "step": 210 }, { "arg_attn_loss": 1.0075690746307373, "arg_loss": 0.9575352072715759, "executor_loss": 4.733499050140381, "init_attn_loss": 0.0003154374426230788, "init_loss": 0.00014578219270333648, "loss": 10.328314781188965, "op_attn_loss": 0.7301731705665588, "op_loss": 0.39882639050483704, "phase": "curriculum_train", "role_loss": 0.22470423579216003, "stage_max_len": 6, "state_loss": 3.331906318664551, "state_train_accuracy": 0.0833333358168602, "step": 280 }, { "arg_attn_loss": 0.5318492650985718, "arg_loss": 0.4931265413761139, "executor_loss": 0.31158560514450073, "init_attn_loss": 0.000763318850658834, "init_loss": 6.9141105996095575e-06, "loss": 5.179892539978027, "op_attn_loss": 0.4329981803894043, "op_loss": 0.28558823466300964, "phase": "curriculum_train", "role_loss": 0.21482081711292267, "stage_max_len": 6, "state_loss": 1.6694793701171875, "state_train_accuracy": 0.3333333432674408, "step": 350 }, { "arg_attn_loss": 0.6200889348983765, "arg_loss": 0.5487199425697327, "executor_loss": 2.2544875144958496, "init_attn_loss": 0.00011264382919762284, "init_loss": 3.218595884391107e-05, "loss": 6.0999650955200195, "op_attn_loss": 0.5651086568832397, "op_loss": 0.21308322250843048, "phase": "curriculum_train", "role_loss": 0.12788312137126923, "stage_max_len": 6, "state_loss": 1.3503047227859497, "state_train_accuracy": 0.7272727489471436, "step": 420 }, { "arg_attn_loss": 0.6377592086791992, "arg_loss": 0.6065531969070435, "executor_loss": 0.4672502279281616, "init_attn_loss": 0.0003474347176961601, "init_loss": 1.8715669284574687e-05, "loss": 6.1048479080200195, "op_attn_loss": 0.513096272945404, "op_loss": 0.20394985377788544, "phase": "curriculum_train", "role_loss": 0.27930837869644165, "stage_max_len": 6, "state_loss": 0.7474269270896912, "state_train_accuracy": 0.75, "step": 490 }, { "arg_attn_loss": 0.35595569014549255, "arg_loss": 0.3529312312602997, "executor_loss": 1.2908166646957397, "init_attn_loss": 9.619701449992135e-05, "init_loss": 5.960462203802308e-07, "loss": 3.7860183715820312, "op_attn_loss": 0.3727604150772095, "op_loss": 0.21929188072681427, "phase": "curriculum_train", "role_loss": 0.040322232991456985, "stage_max_len": 6, "state_loss": 1.196681261062622, "state_train_accuracy": 1.0, "step": 560 }, { "arg_attn_loss": 0.6579608917236328, "arg_loss": 0.62535160779953, "executor_loss": 3.176943302154541, "init_attn_loss": 2.968260378111154e-05, "init_loss": 9.894321920000948e-06, "loss": 6.782617092132568, "op_attn_loss": 0.5155288577079773, "op_loss": 0.27181077003479004, "phase": "curriculum_train", "role_loss": 0.12800464034080505, "stage_max_len": 6, "state_loss": 1.971298336982727, "state_train_accuracy": 0.5, "step": 630 }, { "arg_attn_loss": 0.4417271018028259, "arg_loss": 0.43727967143058777, "executor_loss": 0.901738166809082, "init_attn_loss": 0.00018587702652439475, "init_loss": 2.0861593839072157e-06, "loss": 4.629508018493652, "op_attn_loss": 0.5240081548690796, "op_loss": 0.29447248578071594, "phase": "curriculum_train", "role_loss": 0.03904511779546738, "stage_max_len": 6, "state_loss": 1.508275032043457, "state_train_accuracy": 0.75, "step": 700 }, { "domain_arithmetic_arg_accuracy": 0.90625, "domain_arithmetic_direct_accuracy": 0.0, "domain_arithmetic_domain_arithmetic_direct_accuracy": 0.0, "domain_arithmetic_domain_arithmetic_executor_accuracy": 0.53125, "domain_arithmetic_domain_arithmetic_n": 32.0, "domain_arithmetic_domain_arithmetic_repair_executor_accuracy": 1.0, "domain_arithmetic_executor_accuracy": 0.53125, "domain_arithmetic_init_accuracy": 1.0, "domain_arithmetic_op_accuracy": 0.9895833333333334, "domain_arithmetic_program_exact": 0.53125, "domain_arithmetic_repair_avg_candidates": 339.0, "domain_arithmetic_repair_avg_verified": 1.0, "domain_arithmetic_repair_changed_fraction": 0.46875, "domain_arithmetic_repair_executor_accuracy": 1.0, "domain_arithmetic_repair_found_fraction": 1.0, "domain_arithmetic_repair_program_exact": 0.875, "domain_arithmetic_repair_state_accuracy": 0.9375, "domain_arithmetic_repair_state_prefix_fraction": 0.890625, "domain_arithmetic_state_accuracy": 0.7447916666666666, "domain_arithmetic_state_prefix_fraction": 0.7447916666666666, "domain_boolean_arg_accuracy": 0.9010416666666666, "domain_boolean_direct_accuracy": 0.375, "domain_boolean_domain_boolean_direct_accuracy": 0.375, "domain_boolean_domain_boolean_executor_accuracy": 0.90625, "domain_boolean_domain_boolean_n": 32.0, "domain_boolean_domain_boolean_repair_executor_accuracy": 1.0, "domain_boolean_executor_accuracy": 0.90625, "domain_boolean_init_accuracy": 1.0, "domain_boolean_op_accuracy": 1.0, "domain_boolean_program_exact": 0.46875, "domain_boolean_repair_avg_candidates": 339.0, "domain_boolean_repair_avg_verified": 1.0, "domain_boolean_repair_changed_fraction": 0.09375, "domain_boolean_repair_executor_accuracy": 1.0, "domain_boolean_repair_found_fraction": 1.0, "domain_boolean_repair_program_exact": 0.46875, "domain_boolean_repair_state_accuracy": 0.734375, "domain_boolean_repair_state_prefix_fraction": 0.6458333333333334, "domain_boolean_state_accuracy": 0.71875, "domain_boolean_state_prefix_fraction": 0.6458333333333334, "domain_calendar_arg_accuracy": 0.9166666666666666, "domain_calendar_direct_accuracy": 0.0625, "domain_calendar_domain_calendar_direct_accuracy": 0.0625, "domain_calendar_domain_calendar_executor_accuracy": 0.6875, "domain_calendar_domain_calendar_n": 32.0, "domain_calendar_domain_calendar_repair_executor_accuracy": 1.0, "domain_calendar_executor_accuracy": 0.6875, "domain_calendar_init_accuracy": 1.0, "domain_calendar_op_accuracy": 1.0, "domain_calendar_program_exact": 0.625, "domain_calendar_repair_avg_candidates": 339.0, "domain_calendar_repair_avg_verified": 1.0, "domain_calendar_repair_changed_fraction": 0.3125, "domain_calendar_repair_executor_accuracy": 1.0, "domain_calendar_repair_found_fraction": 1.0, "domain_calendar_repair_program_exact": 0.75, "domain_calendar_repair_state_accuracy": 0.9375, "domain_calendar_repair_state_prefix_fraction": 0.9010416666666666, "domain_calendar_state_accuracy": 0.8489583333333334, "domain_calendar_state_prefix_fraction": 0.8385416666666666, "domain_list_arg_accuracy": 0.9166666666666666, "domain_list_direct_accuracy": 0.03125, "domain_list_domain_list_direct_accuracy": 0.03125, "domain_list_domain_list_executor_accuracy": 0.96875, "domain_list_domain_list_n": 32.0, "domain_list_domain_list_repair_executor_accuracy": 1.0, "domain_list_executor_accuracy": 0.96875, "domain_list_init_accuracy": 1.0, "domain_list_op_accuracy": 1.0, "domain_list_program_exact": 0.5625, "domain_list_repair_avg_candidates": 339.0, "domain_list_repair_avg_verified": 1.0, "domain_list_repair_changed_fraction": 0.03125, "domain_list_repair_executor_accuracy": 1.0, "domain_list_repair_found_fraction": 1.0, "domain_list_repair_program_exact": 0.59375, "domain_list_repair_state_accuracy": 0.96875, "domain_list_repair_state_prefix_fraction": 0.8697916666666666, "domain_list_state_accuracy": 0.9635416666666666, "domain_list_state_prefix_fraction": 0.8645833333333334, "domain_lookup_arg_accuracy": 0.9427083333333334, "domain_lookup_direct_accuracy": 0.0, "domain_lookup_domain_lookup_direct_accuracy": 0.0, "domain_lookup_domain_lookup_executor_accuracy": 0.84375, "domain_lookup_domain_lookup_n": 32.0, "domain_lookup_domain_lookup_repair_executor_accuracy": 1.0, "domain_lookup_executor_accuracy": 0.84375, "domain_lookup_init_accuracy": 1.0, "domain_lookup_op_accuracy": 1.0, "domain_lookup_program_exact": 0.78125, "domain_lookup_repair_avg_candidates": 339.0, "domain_lookup_repair_avg_verified": 1.0, "domain_lookup_repair_changed_fraction": 0.15625, "domain_lookup_repair_executor_accuracy": 1.0, "domain_lookup_repair_found_fraction": 1.0, "domain_lookup_repair_program_exact": 0.78125, "domain_lookup_repair_state_accuracy": 0.90625, "domain_lookup_repair_state_prefix_fraction": 0.8645833333333334, "domain_lookup_state_accuracy": 0.875, "domain_lookup_state_prefix_fraction": 0.859375, "domain_unit_arg_accuracy": 0.9010416666666666, "domain_unit_direct_accuracy": 0.03125, "domain_unit_domain_unit_direct_accuracy": 0.03125, "domain_unit_domain_unit_executor_accuracy": 0.53125, "domain_unit_domain_unit_n": 32.0, "domain_unit_domain_unit_repair_executor_accuracy": 1.0, "domain_unit_executor_accuracy": 0.53125, "domain_unit_init_accuracy": 1.0, "domain_unit_op_accuracy": 0.984375, "domain_unit_program_exact": 0.53125, "domain_unit_repair_avg_candidates": 339.0, "domain_unit_repair_avg_verified": 1.0, "domain_unit_repair_changed_fraction": 0.46875, "domain_unit_repair_executor_accuracy": 1.0, "domain_unit_repair_found_fraction": 1.0, "domain_unit_repair_program_exact": 0.84375, "domain_unit_repair_state_accuracy": 0.921875, "domain_unit_repair_state_prefix_fraction": 0.875, "domain_unit_state_accuracy": 0.8072916666666666, "domain_unit_state_prefix_fraction": 0.8072916666666666, "fresh_paired_mixed_arg_accuracy": 0.9095052083333334, "fresh_paired_mixed_compiler_pair_program_consistency": 0.703125, "fresh_paired_mixed_compiler_pair_state_consistency": 0.7265625, "fresh_paired_mixed_direct_accuracy": 0.109375, "fresh_paired_mixed_domain_arithmetic_direct_accuracy": 0.0, "fresh_paired_mixed_domain_arithmetic_executor_accuracy": 0.6136363636363636, "fresh_paired_mixed_domain_arithmetic_n": 44.0, "fresh_paired_mixed_domain_arithmetic_repair_executor_accuracy": 1.0, "fresh_paired_mixed_domain_boolean_direct_accuracy": 0.42857142857142855, "fresh_paired_mixed_domain_boolean_executor_accuracy": 0.9761904761904762, "fresh_paired_mixed_domain_boolean_n": 42.0, "fresh_paired_mixed_domain_boolean_repair_executor_accuracy": 1.0, "fresh_paired_mixed_domain_calendar_direct_accuracy": 0.18181818181818182, "fresh_paired_mixed_domain_calendar_executor_accuracy": 0.7045454545454546, "fresh_paired_mixed_domain_calendar_n": 44.0, "fresh_paired_mixed_domain_calendar_repair_executor_accuracy": 1.0, "fresh_paired_mixed_domain_list_direct_accuracy": 0.047619047619047616, "fresh_paired_mixed_domain_list_executor_accuracy": 0.7619047619047619, "fresh_paired_mixed_domain_list_n": 42.0, "fresh_paired_mixed_domain_list_repair_executor_accuracy": 1.0, "fresh_paired_mixed_domain_lookup_direct_accuracy": 0.0, "fresh_paired_mixed_domain_lookup_executor_accuracy": 0.7142857142857143, "fresh_paired_mixed_domain_lookup_n": 42.0, "fresh_paired_mixed_domain_lookup_repair_executor_accuracy": 0.9523809523809523, "fresh_paired_mixed_domain_unit_direct_accuracy": 0.0, "fresh_paired_mixed_domain_unit_executor_accuracy": 0.5476190476190477, "fresh_paired_mixed_domain_unit_n": 42.0, "fresh_paired_mixed_domain_unit_repair_executor_accuracy": 0.9761904761904762, "fresh_paired_mixed_executor_accuracy": 0.71875, "fresh_paired_mixed_executor_pair_answer_consistency": 0.7578125, "fresh_paired_mixed_executor_pair_both_correct": 0.625, "fresh_paired_mixed_init_accuracy": 1.0, "fresh_paired_mixed_op_accuracy": 0.9973958333333334, "fresh_paired_mixed_program_exact": 0.55859375, "fresh_paired_mixed_repair_avg_candidates": 339.0, "fresh_paired_mixed_repair_avg_verified": 0.98828125, "fresh_paired_mixed_repair_changed_fraction": 0.26953125, "fresh_paired_mixed_repair_executor_accuracy": 0.98828125, "fresh_paired_mixed_repair_found_fraction": 0.98828125, "fresh_paired_mixed_repair_pair_answer_consistency": 0.984375, "fresh_paired_mixed_repair_pair_both_correct": 0.984375, "fresh_paired_mixed_repair_pair_state_consistency": 0.875, "fresh_paired_mixed_repair_program_exact": 0.67578125, "fresh_paired_mixed_repair_state_accuracy": 0.900390625, "fresh_paired_mixed_repair_state_prefix_fraction": 0.8072916666666666, "fresh_paired_mixed_state_accuracy": 0.8131510416666666, "fresh_paired_mixed_state_prefix_fraction": 0.7727864583333334, "fresh_paraphrase_mixed_arg_accuracy": 0.8723958333333334, "fresh_paraphrase_mixed_direct_accuracy": 0.11979166666666667, "fresh_paraphrase_mixed_domain_arithmetic_direct_accuracy": 0.0, "fresh_paraphrase_mixed_domain_arithmetic_executor_accuracy": 0.28125, "fresh_paraphrase_mixed_domain_arithmetic_n": 32.0, "fresh_paraphrase_mixed_domain_arithmetic_repair_executor_accuracy": 1.0, "fresh_paraphrase_mixed_domain_boolean_direct_accuracy": 0.46875, "fresh_paraphrase_mixed_domain_boolean_executor_accuracy": 0.90625, "fresh_paraphrase_mixed_domain_boolean_n": 32.0, "fresh_paraphrase_mixed_domain_boolean_repair_executor_accuracy": 1.0, "fresh_paraphrase_mixed_domain_calendar_direct_accuracy": 0.25, "fresh_paraphrase_mixed_domain_calendar_executor_accuracy": 0.40625, "fresh_paraphrase_mixed_domain_calendar_n": 32.0, "fresh_paraphrase_mixed_domain_calendar_repair_executor_accuracy": 1.0, "fresh_paraphrase_mixed_domain_list_direct_accuracy": 0.0, "fresh_paraphrase_mixed_domain_list_executor_accuracy": 0.84375, "fresh_paraphrase_mixed_domain_list_n": 32.0, "fresh_paraphrase_mixed_domain_list_repair_executor_accuracy": 1.0, "fresh_paraphrase_mixed_domain_lookup_direct_accuracy": 0.0, "fresh_paraphrase_mixed_domain_lookup_executor_accuracy": 0.78125, "fresh_paraphrase_mixed_domain_lookup_n": 32.0, "fresh_paraphrase_mixed_domain_lookup_repair_executor_accuracy": 0.96875, "fresh_paraphrase_mixed_domain_unit_direct_accuracy": 0.0, "fresh_paraphrase_mixed_domain_unit_executor_accuracy": 0.3125, "fresh_paraphrase_mixed_domain_unit_n": 32.0, "fresh_paraphrase_mixed_domain_unit_repair_executor_accuracy": 0.96875, "fresh_paraphrase_mixed_executor_accuracy": 0.5885416666666666, "fresh_paraphrase_mixed_init_accuracy": 1.0, "fresh_paraphrase_mixed_op_accuracy": 0.9921875, "fresh_paraphrase_mixed_program_exact": 0.4010416666666667, "fresh_paraphrase_mixed_repair_avg_candidates": 339.0, "fresh_paraphrase_mixed_repair_avg_verified": 0.9895833333333334, "fresh_paraphrase_mixed_repair_changed_fraction": 0.4010416666666667, "fresh_paraphrase_mixed_repair_executor_accuracy": 0.9895833333333334, "fresh_paraphrase_mixed_repair_found_fraction": 0.9895833333333334, "fresh_paraphrase_mixed_repair_program_exact": 0.578125, "fresh_paraphrase_mixed_repair_state_accuracy": 0.8498263888888888, "fresh_paraphrase_mixed_repair_state_prefix_fraction": 0.7439236111111112, "fresh_paraphrase_mixed_state_accuracy": 0.734375, "fresh_paraphrase_mixed_state_prefix_fraction": 0.6944444444444444, "fresh_standard_mixed_arg_accuracy": 0.9253472222222222, "fresh_standard_mixed_direct_accuracy": 0.09375, "fresh_standard_mixed_domain_arithmetic_direct_accuracy": 0.0, "fresh_standard_mixed_domain_arithmetic_executor_accuracy": 0.75, "fresh_standard_mixed_domain_arithmetic_n": 32.0, "fresh_standard_mixed_domain_arithmetic_repair_executor_accuracy": 1.0, "fresh_standard_mixed_domain_boolean_direct_accuracy": 0.5, "fresh_standard_mixed_domain_boolean_executor_accuracy": 0.84375, "fresh_standard_mixed_domain_boolean_n": 32.0, "fresh_standard_mixed_domain_boolean_repair_executor_accuracy": 1.0, "fresh_standard_mixed_domain_calendar_direct_accuracy": 0.0625, "fresh_standard_mixed_domain_calendar_executor_accuracy": 0.625, "fresh_standard_mixed_domain_calendar_n": 32.0, "fresh_standard_mixed_domain_calendar_repair_executor_accuracy": 1.0, "fresh_standard_mixed_domain_list_direct_accuracy": 0.0, "fresh_standard_mixed_domain_list_executor_accuracy": 0.75, "fresh_standard_mixed_domain_list_n": 32.0, "fresh_standard_mixed_domain_list_repair_executor_accuracy": 1.0, "fresh_standard_mixed_domain_lookup_direct_accuracy": 0.0, "fresh_standard_mixed_domain_lookup_executor_accuracy": 0.75, "fresh_standard_mixed_domain_lookup_n": 32.0, "fresh_standard_mixed_domain_lookup_repair_executor_accuracy": 1.0, "fresh_standard_mixed_domain_unit_direct_accuracy": 0.0, "fresh_standard_mixed_domain_unit_executor_accuracy": 0.6875, "fresh_standard_mixed_domain_unit_n": 32.0, "fresh_standard_mixed_domain_unit_repair_executor_accuracy": 1.0, "fresh_standard_mixed_executor_accuracy": 0.734375, "fresh_standard_mixed_init_accuracy": 1.0, "fresh_standard_mixed_op_accuracy": 1.0, "fresh_standard_mixed_program_exact": 0.6197916666666666, "fresh_standard_mixed_repair_avg_candidates": 339.0, "fresh_standard_mixed_repair_avg_verified": 1.0, "fresh_standard_mixed_repair_changed_fraction": 0.265625, "fresh_standard_mixed_repair_executor_accuracy": 1.0, "fresh_standard_mixed_repair_found_fraction": 1.0, "fresh_standard_mixed_repair_program_exact": 0.7239583333333334, "fresh_standard_mixed_repair_state_accuracy": 0.8828125, "fresh_standard_mixed_repair_state_prefix_fraction": 0.8177083333333334, "fresh_standard_mixed_state_accuracy": 0.8029513888888888, "fresh_standard_mixed_state_prefix_fraction": 0.7760416666666666, "hard_paraphrase_mixed_arg_accuracy": 0.81640625, "hard_paraphrase_mixed_direct_accuracy": 0.140625, "hard_paraphrase_mixed_domain_arithmetic_direct_accuracy": 0.0, "hard_paraphrase_mixed_domain_arithmetic_executor_accuracy": 0.09375, "hard_paraphrase_mixed_domain_arithmetic_n": 32.0, "hard_paraphrase_mixed_domain_arithmetic_repair_executor_accuracy": 0.96875, "hard_paraphrase_mixed_domain_boolean_direct_accuracy": 0.59375, "hard_paraphrase_mixed_domain_boolean_executor_accuracy": 0.78125, "hard_paraphrase_mixed_domain_boolean_n": 32.0, "hard_paraphrase_mixed_domain_boolean_repair_executor_accuracy": 1.0, "hard_paraphrase_mixed_domain_calendar_direct_accuracy": 0.21875, "hard_paraphrase_mixed_domain_calendar_executor_accuracy": 0.3125, "hard_paraphrase_mixed_domain_calendar_n": 32.0, "hard_paraphrase_mixed_domain_calendar_repair_executor_accuracy": 1.0, "hard_paraphrase_mixed_domain_list_direct_accuracy": 0.0, "hard_paraphrase_mixed_domain_list_executor_accuracy": 0.78125, "hard_paraphrase_mixed_domain_list_n": 32.0, "hard_paraphrase_mixed_domain_list_repair_executor_accuracy": 1.0, "hard_paraphrase_mixed_domain_lookup_direct_accuracy": 0.0, "hard_paraphrase_mixed_domain_lookup_executor_accuracy": 0.5, "hard_paraphrase_mixed_domain_lookup_n": 32.0, "hard_paraphrase_mixed_domain_lookup_repair_executor_accuracy": 0.96875, "hard_paraphrase_mixed_domain_unit_direct_accuracy": 0.03125, "hard_paraphrase_mixed_domain_unit_executor_accuracy": 0.0, "hard_paraphrase_mixed_domain_unit_n": 32.0, "hard_paraphrase_mixed_domain_unit_repair_executor_accuracy": 1.0, "hard_paraphrase_mixed_executor_accuracy": 0.4114583333333333, "hard_paraphrase_mixed_init_accuracy": 1.0, "hard_paraphrase_mixed_op_accuracy": 0.9498697916666666, "hard_paraphrase_mixed_program_exact": 0.203125, "hard_paraphrase_mixed_repair_avg_candidates": 579.0, "hard_paraphrase_mixed_repair_avg_verified": 0.9895833333333334, "hard_paraphrase_mixed_repair_changed_fraction": 0.578125, "hard_paraphrase_mixed_repair_executor_accuracy": 0.9895833333333334, "hard_paraphrase_mixed_repair_found_fraction": 0.9895833333333334, "hard_paraphrase_mixed_repair_program_exact": 0.359375, "hard_paraphrase_mixed_repair_state_accuracy": 0.7141927083333334, "hard_paraphrase_mixed_repair_state_prefix_fraction": 0.6184895833333334, "hard_paraphrase_mixed_state_accuracy": 0.6783854166666666, "hard_paraphrase_mixed_state_prefix_fraction": 0.640625, "hard_standard_mixed_arg_accuracy": 0.8938802083333334, "hard_standard_mixed_direct_accuracy": 0.109375, "hard_standard_mixed_domain_arithmetic_direct_accuracy": 0.03125, "hard_standard_mixed_domain_arithmetic_executor_accuracy": 0.5, "hard_standard_mixed_domain_arithmetic_n": 32.0, "hard_standard_mixed_domain_arithmetic_repair_executor_accuracy": 1.0, "hard_standard_mixed_domain_boolean_direct_accuracy": 0.375, "hard_standard_mixed_domain_boolean_executor_accuracy": 0.71875, "hard_standard_mixed_domain_boolean_n": 32.0, "hard_standard_mixed_domain_boolean_repair_executor_accuracy": 1.0, "hard_standard_mixed_domain_calendar_direct_accuracy": 0.21875, "hard_standard_mixed_domain_calendar_executor_accuracy": 0.53125, "hard_standard_mixed_domain_calendar_n": 32.0, "hard_standard_mixed_domain_calendar_repair_executor_accuracy": 1.0, "hard_standard_mixed_domain_list_direct_accuracy": 0.03125, "hard_standard_mixed_domain_list_executor_accuracy": 0.625, "hard_standard_mixed_domain_list_n": 32.0, "hard_standard_mixed_domain_list_repair_executor_accuracy": 1.0, "hard_standard_mixed_domain_lookup_direct_accuracy": 0.0, "hard_standard_mixed_domain_lookup_executor_accuracy": 0.5, "hard_standard_mixed_domain_lookup_n": 32.0, "hard_standard_mixed_domain_lookup_repair_executor_accuracy": 1.0, "hard_standard_mixed_domain_unit_direct_accuracy": 0.0, "hard_standard_mixed_domain_unit_executor_accuracy": 0.4375, "hard_standard_mixed_domain_unit_n": 32.0, "hard_standard_mixed_domain_unit_repair_executor_accuracy": 1.0, "hard_standard_mixed_executor_accuracy": 0.5520833333333334, "hard_standard_mixed_init_accuracy": 1.0, "hard_standard_mixed_op_accuracy": 0.9889322916666666, "hard_standard_mixed_program_exact": 0.3802083333333333, "hard_standard_mixed_repair_avg_candidates": 579.0, "hard_standard_mixed_repair_avg_verified": 1.0, "hard_standard_mixed_repair_changed_fraction": 0.4479166666666667, "hard_standard_mixed_repair_executor_accuracy": 1.0, "hard_standard_mixed_repair_found_fraction": 1.0, "hard_standard_mixed_repair_program_exact": 0.5625, "hard_standard_mixed_repair_state_accuracy": 0.8313802083333334, "hard_standard_mixed_repair_state_prefix_fraction": 0.7317708333333334, "hard_standard_mixed_state_accuracy": 0.7408854166666666, "hard_standard_mixed_state_prefix_fraction": 0.701171875, "harder_paraphrase_mixed_arg_accuracy": 0.7197916666666667, "harder_paraphrase_mixed_direct_accuracy": 0.09375, "harder_paraphrase_mixed_domain_arithmetic_direct_accuracy": 0.03125, "harder_paraphrase_mixed_domain_arithmetic_executor_accuracy": 0.0, "harder_paraphrase_mixed_domain_arithmetic_n": 32.0, "harder_paraphrase_mixed_domain_arithmetic_repair_executor_accuracy": 1.0, "harder_paraphrase_mixed_domain_boolean_direct_accuracy": 0.3125, "harder_paraphrase_mixed_domain_boolean_executor_accuracy": 0.25, "harder_paraphrase_mixed_domain_boolean_n": 32.0, "harder_paraphrase_mixed_domain_boolean_repair_executor_accuracy": 1.0, "harder_paraphrase_mixed_domain_calendar_direct_accuracy": 0.1875, "harder_paraphrase_mixed_domain_calendar_executor_accuracy": 0.1875, "harder_paraphrase_mixed_domain_calendar_n": 32.0, "harder_paraphrase_mixed_domain_calendar_repair_executor_accuracy": 1.0, "harder_paraphrase_mixed_domain_list_direct_accuracy": 0.0, "harder_paraphrase_mixed_domain_list_executor_accuracy": 0.375, "harder_paraphrase_mixed_domain_list_n": 32.0, "harder_paraphrase_mixed_domain_list_repair_executor_accuracy": 1.0, "harder_paraphrase_mixed_domain_lookup_direct_accuracy": 0.0, "harder_paraphrase_mixed_domain_lookup_executor_accuracy": 0.1875, "harder_paraphrase_mixed_domain_lookup_n": 32.0, "harder_paraphrase_mixed_domain_lookup_repair_executor_accuracy": 0.9375, "harder_paraphrase_mixed_domain_unit_direct_accuracy": 0.03125, "harder_paraphrase_mixed_domain_unit_executor_accuracy": 0.0, "harder_paraphrase_mixed_domain_unit_n": 32.0, "harder_paraphrase_mixed_domain_unit_repair_executor_accuracy": 1.0, "harder_paraphrase_mixed_executor_accuracy": 0.16666666666666666, "harder_paraphrase_mixed_init_accuracy": 1.0, "harder_paraphrase_mixed_op_accuracy": 0.8536458333333333, "harder_paraphrase_mixed_program_exact": 0.03125, "harder_paraphrase_mixed_repair_avg_candidates": 883.0, "harder_paraphrase_mixed_repair_avg_verified": 0.9895833333333334, "harder_paraphrase_mixed_repair_changed_fraction": 0.8229166666666666, "harder_paraphrase_mixed_repair_executor_accuracy": 0.9895833333333334, "harder_paraphrase_mixed_repair_found_fraction": 0.9895833333333334, "harder_paraphrase_mixed_repair_program_exact": 0.08333333333333333, "harder_paraphrase_mixed_repair_state_accuracy": 0.578125, "harder_paraphrase_mixed_repair_state_prefix_fraction": 0.44635416666666666, "harder_paraphrase_mixed_state_accuracy": 0.5640625, "harder_paraphrase_mixed_state_prefix_fraction": 0.5203125, "harder_standard_mixed_arg_accuracy": 0.8270833333333333, "harder_standard_mixed_direct_accuracy": 0.109375, "harder_standard_mixed_domain_arithmetic_direct_accuracy": 0.03125, "harder_standard_mixed_domain_arithmetic_executor_accuracy": 0.03125, "harder_standard_mixed_domain_arithmetic_n": 32.0, "harder_standard_mixed_domain_arithmetic_repair_executor_accuracy": 1.0, "harder_standard_mixed_domain_boolean_direct_accuracy": 0.5, "harder_standard_mixed_domain_boolean_executor_accuracy": 0.375, "harder_standard_mixed_domain_boolean_n": 32.0, "harder_standard_mixed_domain_boolean_repair_executor_accuracy": 1.0, "harder_standard_mixed_domain_calendar_direct_accuracy": 0.09375, "harder_standard_mixed_domain_calendar_executor_accuracy": 0.40625, "harder_standard_mixed_domain_calendar_n": 32.0, "harder_standard_mixed_domain_calendar_repair_executor_accuracy": 1.0, "harder_standard_mixed_domain_list_direct_accuracy": 0.0, "harder_standard_mixed_domain_list_executor_accuracy": 0.5625, "harder_standard_mixed_domain_list_n": 32.0, "harder_standard_mixed_domain_list_repair_executor_accuracy": 1.0, "harder_standard_mixed_domain_lookup_direct_accuracy": 0.0, "harder_standard_mixed_domain_lookup_executor_accuracy": 0.15625, "harder_standard_mixed_domain_lookup_n": 32.0, "harder_standard_mixed_domain_lookup_repair_executor_accuracy": 0.96875, "harder_standard_mixed_domain_unit_direct_accuracy": 0.03125, "harder_standard_mixed_domain_unit_executor_accuracy": 0.28125, "harder_standard_mixed_domain_unit_n": 32.0, "harder_standard_mixed_domain_unit_repair_executor_accuracy": 1.0, "harder_standard_mixed_executor_accuracy": 0.3020833333333333, "harder_standard_mixed_init_accuracy": 1.0, "harder_standard_mixed_op_accuracy": 0.9416666666666667, "harder_standard_mixed_program_exact": 0.140625, "harder_standard_mixed_repair_avg_candidates": 883.0, "harder_standard_mixed_repair_avg_verified": 0.9947916666666666, "harder_standard_mixed_repair_changed_fraction": 0.6927083333333334, "harder_standard_mixed_repair_executor_accuracy": 0.9947916666666666, "harder_standard_mixed_repair_found_fraction": 0.9947916666666666, "harder_standard_mixed_repair_program_exact": 0.2916666666666667, "harder_standard_mixed_repair_state_accuracy": 0.7114583333333333, "harder_standard_mixed_repair_state_prefix_fraction": 0.6, "harder_standard_mixed_state_accuracy": 0.646875, "harder_standard_mixed_state_prefix_fraction": 0.6098958333333333, "phase": "eval", "step": 700, "val_mixed_arg_accuracy": 0.9178240740740741, "val_mixed_direct_accuracy": 0.10416666666666667, "val_mixed_domain_arithmetic_direct_accuracy": 0.0, "val_mixed_domain_arithmetic_executor_accuracy": 0.4583333333333333, "val_mixed_domain_arithmetic_n": 24.0, "val_mixed_domain_arithmetic_repair_executor_accuracy": 1.0, "val_mixed_domain_boolean_direct_accuracy": 0.375, "val_mixed_domain_boolean_executor_accuracy": 0.8333333333333334, "val_mixed_domain_boolean_n": 24.0, "val_mixed_domain_boolean_repair_executor_accuracy": 1.0, "val_mixed_domain_calendar_direct_accuracy": 0.25, "val_mixed_domain_calendar_executor_accuracy": 0.5416666666666666, "val_mixed_domain_calendar_n": 24.0, "val_mixed_domain_calendar_repair_executor_accuracy": 1.0, "val_mixed_domain_list_direct_accuracy": 0.0, "val_mixed_domain_list_executor_accuracy": 0.875, "val_mixed_domain_list_n": 24.0, "val_mixed_domain_list_repair_executor_accuracy": 1.0, "val_mixed_domain_lookup_direct_accuracy": 0.0, "val_mixed_domain_lookup_executor_accuracy": 0.9166666666666666, "val_mixed_domain_lookup_n": 24.0, "val_mixed_domain_lookup_repair_executor_accuracy": 1.0, "val_mixed_domain_unit_direct_accuracy": 0.0, "val_mixed_domain_unit_executor_accuracy": 0.5416666666666666, "val_mixed_domain_unit_n": 24.0, "val_mixed_domain_unit_repair_executor_accuracy": 1.0, "val_mixed_executor_accuracy": 0.6944444444444444, "val_mixed_init_accuracy": 1.0, "val_mixed_op_accuracy": 0.9976851851851852, "val_mixed_program_exact": 0.5833333333333334, "val_mixed_repair_avg_candidates": 339.0, "val_mixed_repair_avg_verified": 1.0, "val_mixed_repair_changed_fraction": 0.3055555555555556, "val_mixed_repair_executor_accuracy": 1.0, "val_mixed_repair_found_fraction": 1.0, "val_mixed_repair_program_exact": 0.7291666666666666, "val_mixed_repair_state_accuracy": 0.9108796296296297, "val_mixed_repair_state_prefix_fraction": 0.8495370370370371, "val_mixed_state_accuracy": 0.8067129629629629, "val_mixed_state_prefix_fraction": 0.7800925925925926 } ] }