{ "metadata": { "python": "3.12.3", "platform": "Linux-6.8.0-62-generic-x86_64-with-glibc2.39", "torch": "2.8.0+cu128", "cuda_available": true, "gpu_name": "NVIDIA RTX 6000 Ada Generation", "args": { "run_name": "main_slot_repair_distill_s512", "compiler_checkpoint": "large_artifacts/qwen_slot_repair_distillation/checkpoints/fixed_compiler_step00800", "train_examples": 512, "val_examples": 128, "eval_examples": 256, "eval_pairs": 256, "train_length": 24, "eval_length": 24, "train_template_modes": "standard,paraphrase", "repair_topk": 3, "repair_max_edits": 2, "repair_max_pair_arg_slots": 24, "qwen_batch_size": 8, "max_length": 384, "editor_d_model": 128, "editor_layers": 3, "editor_heads": 4, "editor_ff_mult": 4, "editor_dropout": 0.05, "editor_epochs": 18, "editor_lr": 0.001, "editor_weight_decay": 0.0001, "editor_target_mode": "oracle_or_gold", "changed_slot_weight": 4.0, "base_copy_logit_bias": 5.0, "edit_gate_loss_weight": 1.0, "edit_value_loss_weight": 1.0, "edit_gate_pos_weight": 8.0, "unchanged_value_loss_weight": 0.05, "edit_threshold_grid": "0.15,0.2,0.25,0.3,0.4,0.5,0.6,0.7", "init_loss_weight": 1.0, "op_loss_weight": 1.0, "arg_loss_weight": 1.0, "grad_clip": 1.0, "seed": 66, "eval_seed": 66001 }, "compiler_args": { "model_id": "Qwen/Qwen3-4B", "modulus": 97, "max_steps": 24, "head_width": 768, "rank_temperature": 1.0, "arg_reader_mode": "after_op", "arg_window": 8, "arg_distance_temperature": 2.0 }, "feature_names": [ "bias", "candidate_prior", "prior_delta", "candidate_prior_per_slot", "prior_delta_per_edit", "base_prior", "edit_count", "changed", "init_edit", "op_edit_count", "arg_edit_count", "edit_fraction", "first_edit_pos_norm", "last_edit_pos_norm", "mean_edit_pos_norm", "edit_span_norm", "mean_new_rank", "max_new_rank", "mean_changed_margin", "min_changed_margin", "mean_changed_entropy", "mean_logp_delta", "min_logp_delta", "max_logp_delta", "base_init_margin", "base_op_margin_mean", "base_op_margin_min", "base_arg_margin_mean", "base_arg_margin_min", "base_op_entropy_mean", "base_arg_entropy_mean", "candidate_answer_norm", "candidate_final_soft_logp", "candidate_state_soft_logp_mean", "candidate_state_soft_logp_min", "base_final_soft_logp", "base_state_soft_logp_mean", "answer_equals_base", "state_same_fraction_as_base", "state_prefix_same_as_base", "mean_old_value_norm", "mean_new_value_norm", "mean_abs_value_delta_norm", "arg_new_value_mean_norm", "arg_abs_delta_mean_norm", "op_new_add_frac", "op_new_sub_frac", "op_new_mul_frac", "candidate_add_frac", "candidate_sub_frac", "candidate_mul_frac", "candidate_arg_mean_norm", "candidate_arg_std_norm", "base_add_frac", "base_sub_frac", "base_mul_frac" ], "trace_feature_names": [ "is_global", "is_step", "active", "step_pos_norm", "candidate_prior", "prior_delta", "edit_count", "init_value_norm", "init_changed", "answer_norm", "final_soft_logp", "base_final_soft_logp", "answer_equals_base", "op_add", "op_sub", "op_mul", "arg_norm", "state_norm", "state_soft_logp", "base_op_add", "base_op_sub", "base_op_mul", "base_arg_norm", "base_state_norm", "base_state_soft_logp", "op_changed", "arg_changed", "state_same_as_base", "op_logp", "arg_logp", "op_logp_delta", "arg_logp_delta", "op_margin", "arg_margin", "op_entropy", "arg_entropy" ], "editor_type": "gated_edit_policy", "best_editor_epoch": 18, "best_editor_threshold": 0.7, "train_seconds": 849.7058529853821 }, "metrics": { "train_len24": { "n": 512.0, "oracle_found_fraction": 0.853515625, "avg_candidates": 1299.0, "avg_positive_candidates": 1.203125, "base_executor_accuracy": 0.275390625, "base_program_exact": 0.275390625, "base_state_prefix_fraction": 0.5504557291666663, "base_changed_fraction": 0.0, "base_avg_edits": 0.0, "base_in_candidate_set_fraction": 1.0, "oracle_executor_accuracy": 0.853515625, "oracle_program_exact": 0.849609375, "oracle_state_prefix_fraction": 0.8811035156249998, "oracle_changed_fraction": 0.578125, "oracle_avg_edits": 0.814453125, "oracle_in_candidate_set_fraction": 1.0, "editor_executor_accuracy": 0.3125, "editor_program_exact": 0.3125, "editor_state_prefix_fraction": 0.5762532552083331, "editor_changed_fraction": 0.669921875, "editor_avg_edits": 1.16015625, "editor_in_candidate_set_fraction": 0.484375, "editor_gate_precision": 0.7152428810720268, "editor_gate_recall": 0.646969696969697, "editor_gate_f1": 0.6793953858392999, "editor_gate_positive_rate": 0.02379623724489796, "editor_gate_target_positive_rate": 0.026307397959183673, "editor_threshold": 0.7, "editor_oracle_gap_recovered": 0.06418918918918919 }, "val_len24": { "n": 128.0, "oracle_found_fraction": 0.8515625, "avg_candidates": 1299.0, "avg_positive_candidates": 1.2890625, "base_executor_accuracy": 0.328125, "base_program_exact": 0.328125, "base_state_prefix_fraction": 0.5657552083333333, "base_changed_fraction": 0.0, "base_avg_edits": 0.0, "base_in_candidate_set_fraction": 1.0, "oracle_executor_accuracy": 0.8515625, "oracle_program_exact": 0.84375, "oracle_state_prefix_fraction": 0.8753255208333336, "oracle_changed_fraction": 0.5234375, "oracle_avg_edits": 0.6953125, "oracle_in_candidate_set_fraction": 1.0, "editor_executor_accuracy": 0.3515625, "editor_program_exact": 0.3515625, "editor_state_prefix_fraction": 0.6201171874999999, "editor_changed_fraction": 0.6640625, "editor_avg_edits": 1.1640625, "editor_in_candidate_set_fraction": 0.484375, "editor_gate_precision": 0.66, "editor_gate_recall": 0.6513157894736842, "editor_gate_f1": 0.6556291390728477, "editor_gate_positive_rate": 0.02391581632653061, "editor_gate_target_positive_rate": 0.02423469387755102, "editor_threshold": 0.7, "editor_oracle_gap_recovered": 0.04477611940298507 }, "fresh_standard_len24": { "n": 256.0, "oracle_found_fraction": 0.87109375, "avg_candidates": 1299.0, "avg_positive_candidates": 1.40625, "base_executor_accuracy": 0.28515625, "base_program_exact": 0.28515625, "base_state_prefix_fraction": 0.5776367187500001, "base_changed_fraction": 0.0, "base_avg_edits": 0.0, "base_in_candidate_set_fraction": 1.0, "oracle_executor_accuracy": 0.87109375, "oracle_program_exact": 0.87109375, "oracle_state_prefix_fraction": 0.8994140624999998, "oracle_changed_fraction": 0.5859375, "oracle_avg_edits": 0.78125, "oracle_in_candidate_set_fraction": 1.0, "editor_executor_accuracy": 0.27734375, "editor_program_exact": 0.27734375, "editor_state_prefix_fraction": 0.5424804687500001, "editor_changed_fraction": 0.609375, "editor_avg_edits": 0.9765625, "editor_in_candidate_set_fraction": 0.49609375, "editor_gate_precision": 0.6186770428015564, "editor_gate_recall": 0.5162337662337663, "editor_gate_f1": 0.5628318584070796, "editor_gate_positive_rate": 0.020487882653061226, "editor_gate_target_positive_rate": 0.024553571428571428, "editor_threshold": 0.7, "editor_oracle_gap_recovered": -0.013333333333333334 }, "fresh_paraphrase_len24": { "n": 256.0, "oracle_found_fraction": 0.8515625, "avg_candidates": 1299.0, "avg_positive_candidates": 1.27734375, "base_executor_accuracy": 0.25, "base_program_exact": 0.25, "base_state_prefix_fraction": 0.5172526041666671, "base_changed_fraction": 0.0, "base_avg_edits": 0.0, "base_in_candidate_set_fraction": 1.0, "oracle_executor_accuracy": 0.8515625, "oracle_program_exact": 0.8515625, "oracle_state_prefix_fraction": 0.8767903645833334, "oracle_changed_fraction": 0.6015625, "oracle_avg_edits": 0.82421875, "oracle_in_candidate_set_fraction": 1.0, "editor_executor_accuracy": 0.1875, "editor_program_exact": 0.1875, "editor_state_prefix_fraction": 0.4742838541666669, "editor_changed_fraction": 0.75390625, "editor_avg_edits": 1.3359375, "editor_in_candidate_set_fraction": 0.34765625, "editor_gate_precision": 0.6271676300578035, "editor_gate_recall": 0.638235294117647, "editor_gate_f1": 0.6326530612244897, "editor_gate_positive_rate": 0.027582908163265307, "editor_gate_target_positive_rate": 0.027104591836734693, "editor_threshold": 0.7, "editor_oracle_gap_recovered": -0.1038961038961039 }, "fresh_paired_len24": { "n": 512.0, "oracle_found_fraction": 0.861328125, "avg_candidates": 1299.0, "avg_positive_candidates": 1.201171875, "base_executor_accuracy": 0.25390625, "base_program_exact": 0.25390625, "base_state_prefix_fraction": 0.5228678385416667, "base_changed_fraction": 0.0, "base_avg_edits": 0.0, "base_in_candidate_set_fraction": 1.0, "oracle_executor_accuracy": 0.861328125, "oracle_program_exact": 0.857421875, "oracle_state_prefix_fraction": 0.8868815104166666, "oracle_changed_fraction": 0.607421875, "oracle_avg_edits": 0.859375, "oracle_in_candidate_set_fraction": 1.0, "editor_executor_accuracy": 0.248046875, "editor_program_exact": 0.248046875, "editor_state_prefix_fraction": 0.533935546875, "editor_changed_fraction": 0.6796875, "editor_avg_edits": 1.169921875, "editor_in_candidate_set_fraction": 0.439453125, "editor_gate_precision": 0.6694214876033058, "editor_gate_recall": 0.5973451327433629, "editor_gate_f1": 0.6313328137178488, "editor_gate_positive_rate": 0.024115114795918366, "editor_gate_target_positive_rate": 0.02702487244897959, "editor_threshold": 0.7, "editor_oracle_gap_recovered": -0.00964630225080386, "base_pair_answer_consistency": 0.6953125, "base_pair_both_correct": 0.234375, "base_pair_program_consistency": 0.6953125, "base_pair_state_consistency": 0.6953125, "oracle_pair_answer_consistency": 0.8828125, "oracle_pair_both_correct": 0.83984375, "oracle_pair_program_consistency": 0.8828125, "oracle_pair_state_consistency": 0.8828125, "editor_pair_answer_consistency": 0.37890625, "editor_pair_both_correct": 0.16796875, "editor_pair_program_consistency": 0.37109375, "editor_pair_state_consistency": 0.37109375 } }, "train_log": [ { "epoch": 1, "train_loss": 7.917579598724842, "train_gate_loss": 0.623155681991193, "train_value_loss": 7.29442392103374, "train_init_gate_loss": 0.0010433578863739967, "train_op_gate_loss": 0.015596376623442154, "train_arg_gate_loss": 0.606515947576554, "train_init_value_loss": 4.685488092713058, "train_op_value_loss": 0.05353772036615112, "train_arg_value_loss": 2.5553980944678187, "train_target_changed_slots": 1.2890625, "train_target_changed_fraction": 0.026307397959183663, "train_target_oracle_fraction": 0.853515625, "train_target_gold_fallback_fraction": 0.146484375, "train_target_base_fallback_fraction": 0.0, "val_selected_threshold": 0.7, "val_n": 128.0, "val_oracle_found_fraction": 0.8515625, "val_avg_candidates": 1299.0, "val_avg_positive_candidates": 1.2890625, "val_base_executor_accuracy": 0.328125, "val_base_program_exact": 0.328125, "val_base_state_prefix_fraction": 0.5657552083333333, "val_base_changed_fraction": 0.0, "val_base_avg_edits": 0.0, "val_base_in_candidate_set_fraction": 1.0, "val_oracle_executor_accuracy": 0.8515625, "val_oracle_program_exact": 0.84375, "val_oracle_state_prefix_fraction": 0.8753255208333336, "val_oracle_changed_fraction": 0.5234375, "val_oracle_avg_edits": 0.6953125, "val_oracle_in_candidate_set_fraction": 1.0, "val_editor_executor_accuracy": 0.1328125, "val_editor_program_exact": 0.1328125, "val_editor_state_prefix_fraction": 0.439453125, "val_editor_changed_fraction": 0.859375, "val_editor_avg_edits": 1.984375, "val_editor_in_candidate_set_fraction": 0.25, "val_editor_gate_precision": 0.37254901960784315, "val_editor_gate_recall": 0.625, "val_editor_gate_f1": 0.46683046683046686, "val_editor_gate_positive_rate": 0.04065688775510204, "val_editor_gate_target_positive_rate": 0.02423469387755102, "val_editor_threshold": 0.7, "val_editor_oracle_gap_recovered": -0.373134328358209 }, { "epoch": 2, "train_loss": 7.380760028026998, "train_gate_loss": 0.524428570191958, "train_value_loss": 6.856331456452608, "train_init_gate_loss": 3.824010491371155e-06, "train_op_gate_loss": 0.008918674701291707, "train_arg_gate_loss": 0.515506071242271, "train_init_value_loss": 4.561483504250646, "train_op_value_loss": 0.026845045260285083, "train_arg_value_loss": 2.268002894008532, "train_target_changed_slots": 1.2890625, "train_target_changed_fraction": 0.0263073979591837, "train_target_oracle_fraction": 0.853515625, "train_target_gold_fallback_fraction": 0.146484375, "train_target_base_fallback_fraction": 0.0, "val_selected_threshold": 0.7, "val_n": 128.0, "val_oracle_found_fraction": 0.8515625, "val_avg_candidates": 1299.0, "val_avg_positive_candidates": 1.2890625, "val_base_executor_accuracy": 0.328125, "val_base_program_exact": 0.328125, "val_base_state_prefix_fraction": 0.5657552083333333, "val_base_changed_fraction": 0.0, "val_base_avg_edits": 0.0, "val_base_in_candidate_set_fraction": 1.0, "val_oracle_executor_accuracy": 0.8515625, "val_oracle_program_exact": 0.84375, "val_oracle_state_prefix_fraction": 0.8753255208333336, "val_oracle_changed_fraction": 0.5234375, "val_oracle_avg_edits": 0.6953125, "val_oracle_in_candidate_set_fraction": 1.0, "val_editor_executor_accuracy": 0.265625, "val_editor_program_exact": 0.265625, "val_editor_state_prefix_fraction": 0.5553385416666665, "val_editor_changed_fraction": 0.671875, "val_editor_avg_edits": 1.15625, "val_editor_in_candidate_set_fraction": 0.4296875, "val_editor_gate_precision": 0.6026490066225165, "val_editor_gate_recall": 0.5986842105263158, "val_editor_gate_f1": 0.6006600660066007, "val_editor_gate_positive_rate": 0.024075255102040817, "val_editor_gate_target_positive_rate": 0.02423469387755102, "val_editor_threshold": 0.7, "val_editor_oracle_gap_recovered": -0.11940298507462686 }, { "epoch": 3, "train_loss": 6.90853556804359, "train_gate_loss": 0.5063149038178381, "train_value_loss": 6.402220665477216, "train_init_gate_loss": 3.46451997756958e-06, "train_op_gate_loss": 0.009572886374300538, "train_arg_gate_loss": 0.4967385526833823, "train_init_value_loss": 4.228447898523882, "train_op_value_loss": 0.04100410283329747, "train_arg_value_loss": 2.13276866520755, "train_target_changed_slots": 1.2890625, "train_target_changed_fraction": 0.02630739795918371, "train_target_oracle_fraction": 0.853515625, "train_target_gold_fallback_fraction": 0.146484375, "train_target_base_fallback_fraction": 0.0, "val_selected_threshold": 0.7, "val_n": 128.0, "val_oracle_found_fraction": 0.8515625, "val_avg_candidates": 1299.0, "val_avg_positive_candidates": 1.2890625, "val_base_executor_accuracy": 0.328125, "val_base_program_exact": 0.328125, "val_base_state_prefix_fraction": 0.5657552083333333, "val_base_changed_fraction": 0.0, "val_base_avg_edits": 0.0, "val_base_in_candidate_set_fraction": 1.0, "val_oracle_executor_accuracy": 0.8515625, "val_oracle_program_exact": 0.84375, "val_oracle_state_prefix_fraction": 0.8753255208333336, "val_oracle_changed_fraction": 0.5234375, "val_oracle_avg_edits": 0.6953125, "val_oracle_in_candidate_set_fraction": 1.0, "val_editor_executor_accuracy": 0.2265625, "val_editor_program_exact": 0.2265625, "val_editor_state_prefix_fraction": 0.5, "val_editor_changed_fraction": 0.78125, "val_editor_avg_edits": 1.5390625, "val_editor_in_candidate_set_fraction": 0.328125, "val_editor_gate_precision": 0.51010101010101, "val_editor_gate_recall": 0.6644736842105263, "val_editor_gate_f1": 0.5771428571428571, "val_editor_gate_positive_rate": 0.031568877551020405, "val_editor_gate_target_positive_rate": 0.02423469387755102, "val_editor_threshold": 0.7, "val_editor_oracle_gap_recovered": -0.19402985074626866 }, { "epoch": 4, "train_loss": 6.482887657824904, "train_gate_loss": 0.47901242963416735, "train_value_loss": 6.0038752243854105, "train_init_gate_loss": 6.928108632564545e-06, "train_op_gate_loss": 0.007788796833808398, "train_arg_gate_loss": 0.4712167047400726, "train_init_value_loss": 3.947946386411786, "train_op_value_loss": 0.028890581054728415, "train_arg_value_loss": 2.027038248372264, "train_target_changed_slots": 1.2890625, "train_target_changed_fraction": 0.026307397959183708, "train_target_oracle_fraction": 0.853515625, "train_target_gold_fallback_fraction": 0.146484375, "train_target_base_fallback_fraction": 0.0, "val_selected_threshold": 0.7, "val_n": 128.0, "val_oracle_found_fraction": 0.8515625, "val_avg_candidates": 1299.0, "val_avg_positive_candidates": 1.2890625, "val_base_executor_accuracy": 0.328125, "val_base_program_exact": 0.328125, "val_base_state_prefix_fraction": 0.5657552083333333, "val_base_changed_fraction": 0.0, "val_base_avg_edits": 0.0, "val_base_in_candidate_set_fraction": 1.0, "val_oracle_executor_accuracy": 0.8515625, "val_oracle_program_exact": 0.84375, "val_oracle_state_prefix_fraction": 0.8753255208333336, "val_oracle_changed_fraction": 0.5234375, "val_oracle_avg_edits": 0.6953125, "val_oracle_in_candidate_set_fraction": 1.0, "val_editor_executor_accuracy": 0.28125, "val_editor_program_exact": 0.28125, "val_editor_state_prefix_fraction": 0.5722656249999998, "val_editor_changed_fraction": 0.6875, "val_editor_avg_edits": 1.140625, "val_editor_in_candidate_set_fraction": 0.421875, "val_editor_gate_precision": 0.6530612244897959, "val_editor_gate_recall": 0.631578947368421, "val_editor_gate_f1": 0.6421404682274247, "val_editor_gate_positive_rate": 0.0234375, "val_editor_gate_target_positive_rate": 0.02423469387755102, "val_editor_threshold": 0.7, "val_editor_oracle_gap_recovered": -0.08955223880597014 }, { "epoch": 5, "train_loss": 6.132786842063069, "train_gate_loss": 0.474977378755284, "train_value_loss": 5.6578094800934196, "train_init_gate_loss": 1.216307282447815e-06, "train_op_gate_loss": 0.006754979667512817, "train_arg_gate_loss": 0.4682211828403524, "train_init_value_loss": 3.6749405276495963, "train_op_value_loss": 0.017818995539074622, "train_arg_value_loss": 1.9650499490089715, "train_target_changed_slots": 1.2890625, "train_target_changed_fraction": 0.02630739795918368, "train_target_oracle_fraction": 0.853515625, "train_target_gold_fallback_fraction": 0.146484375, "train_target_base_fallback_fraction": 0.0, "val_selected_threshold": 0.7, "val_n": 128.0, "val_oracle_found_fraction": 0.8515625, "val_avg_candidates": 1299.0, "val_avg_positive_candidates": 1.2890625, "val_base_executor_accuracy": 0.328125, "val_base_program_exact": 0.328125, "val_base_state_prefix_fraction": 0.5657552083333333, "val_base_changed_fraction": 0.0, "val_base_avg_edits": 0.0, "val_base_in_candidate_set_fraction": 1.0, "val_oracle_executor_accuracy": 0.8515625, "val_oracle_program_exact": 0.84375, "val_oracle_state_prefix_fraction": 0.8753255208333336, "val_oracle_changed_fraction": 0.5234375, "val_oracle_avg_edits": 0.6953125, "val_oracle_in_candidate_set_fraction": 1.0, "val_editor_executor_accuracy": 0.3125, "val_editor_program_exact": 0.3125, "val_editor_state_prefix_fraction": 0.5507812499999998, "val_editor_changed_fraction": 0.59375, "val_editor_avg_edits": 0.984375, "val_editor_in_candidate_set_fraction": 0.5078125, "val_editor_gate_precision": 0.6141732283464567, "val_editor_gate_recall": 0.5131578947368421, "val_editor_gate_f1": 0.5591397849462365, "val_editor_gate_positive_rate": 0.02024872448979592, "val_editor_gate_target_positive_rate": 0.02423469387755102, "val_editor_threshold": 0.7, "val_editor_oracle_gap_recovered": -0.029850746268656716 }, { "epoch": 6, "train_loss": 5.9173253062181175, "train_gate_loss": 0.47256252652732655, "train_value_loss": 5.444762779865414, "train_init_gate_loss": 1.3541430234909058e-06, "train_op_gate_loss": 0.0054824893983654865, "train_arg_gate_loss": 0.46707868421799503, "train_init_value_loss": 3.5010761548765004, "train_op_value_loss": 0.010406345664574379, "train_arg_value_loss": 1.933280297089368, "train_target_changed_slots": 1.2890625, "train_target_changed_fraction": 0.02630739795918369, "train_target_oracle_fraction": 0.853515625, "train_target_gold_fallback_fraction": 0.146484375, "train_target_base_fallback_fraction": 0.0, "val_selected_threshold": 0.7, "val_n": 128.0, "val_oracle_found_fraction": 0.8515625, "val_avg_candidates": 1299.0, "val_avg_positive_candidates": 1.2890625, "val_base_executor_accuracy": 0.328125, "val_base_program_exact": 0.328125, "val_base_state_prefix_fraction": 0.5657552083333333, "val_base_changed_fraction": 0.0, "val_base_avg_edits": 0.0, "val_base_in_candidate_set_fraction": 1.0, "val_oracle_executor_accuracy": 0.8515625, "val_oracle_program_exact": 0.84375, "val_oracle_state_prefix_fraction": 0.8753255208333336, "val_oracle_changed_fraction": 0.5234375, "val_oracle_avg_edits": 0.6953125, "val_oracle_in_candidate_set_fraction": 1.0, "val_editor_executor_accuracy": 0.296875, "val_editor_program_exact": 0.296875, "val_editor_state_prefix_fraction": 0.5751953124999998, "val_editor_changed_fraction": 0.625, "val_editor_avg_edits": 0.9609375, "val_editor_in_candidate_set_fraction": 0.46875, "val_editor_gate_precision": 0.7096774193548387, "val_editor_gate_recall": 0.5789473684210527, "val_editor_gate_f1": 0.6376811594202899, "val_editor_gate_positive_rate": 0.019770408163265307, "val_editor_gate_target_positive_rate": 0.02423469387755102, "val_editor_threshold": 0.7, "val_editor_oracle_gap_recovered": -0.05970149253731343 }, { "epoch": 7, "train_loss": 5.732467763591558, "train_gate_loss": 0.4519833525191643, "train_value_loss": 5.2804844179190695, "train_init_gate_loss": 1.7285346984863281e-06, "train_op_gate_loss": 0.006010107254254327, "train_arg_gate_loss": 0.44597151710331673, "train_init_value_loss": 3.384405711432919, "train_op_value_loss": 0.012671807826122067, "train_arg_value_loss": 1.8834068914875388, "train_target_changed_slots": 1.2890625, "train_target_changed_fraction": 0.026307397959183684, "train_target_oracle_fraction": 0.853515625, "train_target_gold_fallback_fraction": 0.146484375, "train_target_base_fallback_fraction": 0.0, "val_selected_threshold": 0.7, "val_n": 128.0, "val_oracle_found_fraction": 0.8515625, "val_avg_candidates": 1299.0, "val_avg_positive_candidates": 1.2890625, "val_base_executor_accuracy": 0.328125, "val_base_program_exact": 0.328125, "val_base_state_prefix_fraction": 0.5657552083333333, "val_base_changed_fraction": 0.0, "val_base_avg_edits": 0.0, "val_base_in_candidate_set_fraction": 1.0, "val_oracle_executor_accuracy": 0.8515625, "val_oracle_program_exact": 0.84375, "val_oracle_state_prefix_fraction": 0.8753255208333336, "val_oracle_changed_fraction": 0.5234375, "val_oracle_avg_edits": 0.6953125, "val_oracle_in_candidate_set_fraction": 1.0, "val_editor_executor_accuracy": 0.296875, "val_editor_program_exact": 0.296875, "val_editor_state_prefix_fraction": 0.5589192708333333, "val_editor_changed_fraction": 0.6640625, "val_editor_avg_edits": 1.0859375, "val_editor_in_candidate_set_fraction": 0.4375, "val_editor_gate_precision": 0.6619718309859155, "val_editor_gate_recall": 0.618421052631579, "val_editor_gate_f1": 0.6394557823129252, "val_editor_gate_positive_rate": 0.02264030612244898, "val_editor_gate_target_positive_rate": 0.02423469387755102, "val_editor_threshold": 0.7, "val_editor_oracle_gap_recovered": -0.05970149253731343 }, { "epoch": 8, "train_loss": 5.569198497571051, "train_gate_loss": 0.45405629423476057, "train_value_loss": 5.115142202470452, "train_init_gate_loss": 1.0523945093154907e-06, "train_op_gate_loss": 0.008020361265329845, "train_arg_gate_loss": 0.4460348798602354, "train_init_value_loss": 3.2544414510484785, "train_op_value_loss": 0.0171209153634721, "train_arg_value_loss": 1.843579841661267, "train_target_changed_slots": 1.2890625, "train_target_changed_fraction": 0.026307397959183704, "train_target_oracle_fraction": 0.853515625, "train_target_gold_fallback_fraction": 0.146484375, "train_target_base_fallback_fraction": 0.0, "val_selected_threshold": 0.7, "val_n": 128.0, "val_oracle_found_fraction": 0.8515625, "val_avg_candidates": 1299.0, "val_avg_positive_candidates": 1.2890625, "val_base_executor_accuracy": 0.328125, "val_base_program_exact": 0.328125, "val_base_state_prefix_fraction": 0.5657552083333333, "val_base_changed_fraction": 0.0, "val_base_avg_edits": 0.0, "val_base_in_candidate_set_fraction": 1.0, "val_oracle_executor_accuracy": 0.8515625, "val_oracle_program_exact": 0.84375, "val_oracle_state_prefix_fraction": 0.8753255208333336, "val_oracle_changed_fraction": 0.5234375, "val_oracle_avg_edits": 0.6953125, "val_oracle_in_candidate_set_fraction": 1.0, "val_editor_executor_accuracy": 0.3046875, "val_editor_program_exact": 0.3046875, "val_editor_state_prefix_fraction": 0.5621744791666667, "val_editor_changed_fraction": 0.625, "val_editor_avg_edits": 1.0234375, "val_editor_in_candidate_set_fraction": 0.4609375, "val_editor_gate_precision": 0.6564885496183206, "val_editor_gate_recall": 0.5657894736842105, "val_editor_gate_f1": 0.607773851590106, "val_editor_gate_positive_rate": 0.020886479591836735, "val_editor_gate_target_positive_rate": 0.02423469387755102, "val_editor_threshold": 0.7, "val_editor_oracle_gap_recovered": -0.04477611940298507 }, { "epoch": 9, "train_loss": 5.494235629681498, "train_gate_loss": 0.43729284381697653, "train_value_loss": 5.056942788418382, "train_init_gate_loss": 6.51925802230835e-07, "train_op_gate_loss": 0.004602778452639941, "train_arg_gate_loss": 0.4326894138357602, "train_init_value_loss": 3.210622046957724, "train_op_value_loss": 0.016582611055007856, "train_arg_value_loss": 1.8297381332376972, "train_target_changed_slots": 1.2890625, "train_target_changed_fraction": 0.026307397959183694, "train_target_oracle_fraction": 0.853515625, "train_target_gold_fallback_fraction": 0.146484375, "train_target_base_fallback_fraction": 0.0, "val_selected_threshold": 0.7, "val_n": 128.0, "val_oracle_found_fraction": 0.8515625, "val_avg_candidates": 1299.0, "val_avg_positive_candidates": 1.2890625, "val_base_executor_accuracy": 0.328125, "val_base_program_exact": 0.328125, "val_base_state_prefix_fraction": 0.5657552083333333, "val_base_changed_fraction": 0.0, "val_base_avg_edits": 0.0, "val_base_in_candidate_set_fraction": 1.0, "val_oracle_executor_accuracy": 0.8515625, "val_oracle_program_exact": 0.84375, "val_oracle_state_prefix_fraction": 0.8753255208333336, "val_oracle_changed_fraction": 0.5234375, "val_oracle_avg_edits": 0.6953125, "val_oracle_in_candidate_set_fraction": 1.0, "val_editor_executor_accuracy": 0.2734375, "val_editor_program_exact": 0.2734375, "val_editor_state_prefix_fraction": 0.5602213541666665, "val_editor_changed_fraction": 0.671875, "val_editor_avg_edits": 1.25, "val_editor_in_candidate_set_fraction": 0.4375, "val_editor_gate_precision": 0.5900621118012422, "val_editor_gate_recall": 0.625, "val_editor_gate_f1": 0.6070287539936102, "val_editor_gate_positive_rate": 0.025669642857142856, "val_editor_gate_target_positive_rate": 0.02423469387755102, "val_editor_threshold": 0.7, "val_editor_oracle_gap_recovered": -0.1044776119402985 }, { "epoch": 10, "train_loss": 5.3366265702061355, "train_gate_loss": 0.42282128956139786, "train_value_loss": 4.913805280812085, "train_init_gate_loss": 1.043081283569336e-07, "train_op_gate_loss": 0.006832935305216982, "train_arg_gate_loss": 0.41598825049732113, "train_init_value_loss": 3.116850161459297, "train_op_value_loss": 0.01449032768297734, "train_arg_value_loss": 1.7824647831730545, "train_target_changed_slots": 1.2890625, "train_target_changed_fraction": 0.02630739795918369, "train_target_oracle_fraction": 0.853515625, "train_target_gold_fallback_fraction": 0.146484375, "train_target_base_fallback_fraction": 0.0, "val_selected_threshold": 0.7, "val_n": 128.0, "val_oracle_found_fraction": 0.8515625, "val_avg_candidates": 1299.0, "val_avg_positive_candidates": 1.2890625, "val_base_executor_accuracy": 0.328125, "val_base_program_exact": 0.328125, "val_base_state_prefix_fraction": 0.5657552083333333, "val_base_changed_fraction": 0.0, "val_base_avg_edits": 0.0, "val_base_in_candidate_set_fraction": 1.0, "val_oracle_executor_accuracy": 0.8515625, "val_oracle_program_exact": 0.84375, "val_oracle_state_prefix_fraction": 0.8753255208333336, "val_oracle_changed_fraction": 0.5234375, "val_oracle_avg_edits": 0.6953125, "val_oracle_in_candidate_set_fraction": 1.0, "val_editor_executor_accuracy": 0.2265625, "val_editor_program_exact": 0.2265625, "val_editor_state_prefix_fraction": 0.5250651041666667, "val_editor_changed_fraction": 0.7421875, "val_editor_avg_edits": 1.46875, "val_editor_in_candidate_set_fraction": 0.3359375, "val_editor_gate_precision": 0.5103092783505154, "val_editor_gate_recall": 0.6513157894736842, "val_editor_gate_f1": 0.5722543352601156, "val_editor_gate_positive_rate": 0.03093112244897959, "val_editor_gate_target_positive_rate": 0.02423469387755102, "val_editor_threshold": 0.7, "val_editor_oracle_gap_recovered": -0.19402985074626866 }, { "epoch": 11, "train_loss": 5.350213652011007, "train_gate_loss": 0.4320841034168552, "train_value_loss": 4.918129556812346, "train_init_gate_loss": 6.705522537231445e-08, "train_op_gate_loss": 0.0063776820173971766, "train_arg_gate_loss": 0.42570635511219734, "train_init_value_loss": 3.1253434200771153, "train_op_value_loss": 0.012690204123037245, "train_arg_value_loss": 1.7800959349842742, "train_target_changed_slots": 1.2890625, "train_target_changed_fraction": 0.026307397959183666, "train_target_oracle_fraction": 0.853515625, "train_target_gold_fallback_fraction": 0.146484375, "train_target_base_fallback_fraction": 0.0, "val_selected_threshold": 0.7, "val_n": 128.0, "val_oracle_found_fraction": 0.8515625, "val_avg_candidates": 1299.0, "val_avg_positive_candidates": 1.2890625, "val_base_executor_accuracy": 0.328125, "val_base_program_exact": 0.328125, "val_base_state_prefix_fraction": 0.5657552083333333, "val_base_changed_fraction": 0.0, "val_base_avg_edits": 0.0, "val_base_in_candidate_set_fraction": 1.0, "val_oracle_executor_accuracy": 0.8515625, "val_oracle_program_exact": 0.84375, "val_oracle_state_prefix_fraction": 0.8753255208333336, "val_oracle_changed_fraction": 0.5234375, "val_oracle_avg_edits": 0.6953125, "val_oracle_in_candidate_set_fraction": 1.0, "val_editor_executor_accuracy": 0.3046875, "val_editor_program_exact": 0.3046875, "val_editor_state_prefix_fraction": 0.57421875, "val_editor_changed_fraction": 0.6796875, "val_editor_avg_edits": 1.2421875, "val_editor_in_candidate_set_fraction": 0.4140625, "val_editor_gate_precision": 0.6037735849056604, "val_editor_gate_recall": 0.631578947368421, "val_editor_gate_f1": 0.6173633440514469, "val_editor_gate_positive_rate": 0.025350765306122448, "val_editor_gate_target_positive_rate": 0.02423469387755102, "val_editor_threshold": 0.7, "val_editor_oracle_gap_recovered": -0.04477611940298507 }, { "epoch": 12, "train_loss": 5.117988193407655, "train_gate_loss": 0.4147532347633387, "train_value_loss": 4.70323495939374, "train_init_gate_loss": 9.685754776000977e-08, "train_op_gate_loss": 0.005738961374239807, "train_arg_gate_loss": 0.40901417693385156, "train_init_value_loss": 2.96352767967619, "train_op_value_loss": 0.0076164571654946615, "train_arg_value_loss": 1.7320908232359216, "train_target_changed_slots": 1.2890625, "train_target_changed_fraction": 0.026307397959183715, "train_target_oracle_fraction": 0.853515625, "train_target_gold_fallback_fraction": 0.146484375, "train_target_base_fallback_fraction": 0.0, "val_selected_threshold": 0.7, "val_n": 128.0, "val_oracle_found_fraction": 0.8515625, "val_avg_candidates": 1299.0, "val_avg_positive_candidates": 1.2890625, "val_base_executor_accuracy": 0.328125, "val_base_program_exact": 0.328125, "val_base_state_prefix_fraction": 0.5657552083333333, "val_base_changed_fraction": 0.0, "val_base_avg_edits": 0.0, "val_base_in_candidate_set_fraction": 1.0, "val_oracle_executor_accuracy": 0.8515625, "val_oracle_program_exact": 0.84375, "val_oracle_state_prefix_fraction": 0.8753255208333336, "val_oracle_changed_fraction": 0.5234375, "val_oracle_avg_edits": 0.6953125, "val_oracle_in_candidate_set_fraction": 1.0, "val_editor_executor_accuracy": 0.1875, "val_editor_program_exact": 0.1875, "val_editor_state_prefix_fraction": 0.43261718750000006, "val_editor_changed_fraction": 0.796875, "val_editor_avg_edits": 1.84375, "val_editor_in_candidate_set_fraction": 0.25, "val_editor_gate_precision": 0.4533898305084746, "val_editor_gate_recall": 0.7039473684210527, "val_editor_gate_f1": 0.5515463917525774, "val_editor_gate_positive_rate": 0.03762755102040816, "val_editor_gate_target_positive_rate": 0.02423469387755102, "val_editor_threshold": 0.7, "val_editor_oracle_gap_recovered": -0.26865671641791045 }, { "epoch": 13, "train_loss": 5.1369093612302095, "train_gate_loss": 0.4137288981837628, "train_value_loss": 4.723180464934558, "train_init_gate_loss": 1.2479722499847412e-07, "train_op_gate_loss": 0.008551938014110458, "train_arg_gate_loss": 0.40517683451616904, "train_init_value_loss": 2.9746805321774445, "train_op_value_loss": 0.024816717476022366, "train_arg_value_loss": 1.7236832081107423, "train_target_changed_slots": 1.2890625, "train_target_changed_fraction": 0.026307397959183708, "train_target_oracle_fraction": 0.853515625, "train_target_gold_fallback_fraction": 0.146484375, "train_target_base_fallback_fraction": 0.0, "val_selected_threshold": 0.7, "val_n": 128.0, "val_oracle_found_fraction": 0.8515625, "val_avg_candidates": 1299.0, "val_avg_positive_candidates": 1.2890625, "val_base_executor_accuracy": 0.328125, "val_base_program_exact": 0.328125, "val_base_state_prefix_fraction": 0.5657552083333333, "val_base_changed_fraction": 0.0, "val_base_avg_edits": 0.0, "val_base_in_candidate_set_fraction": 1.0, "val_oracle_executor_accuracy": 0.8515625, "val_oracle_program_exact": 0.84375, "val_oracle_state_prefix_fraction": 0.8753255208333336, "val_oracle_changed_fraction": 0.5234375, "val_oracle_avg_edits": 0.6953125, "val_oracle_in_candidate_set_fraction": 1.0, "val_editor_executor_accuracy": 0.15625, "val_editor_program_exact": 0.15625, "val_editor_state_prefix_fraction": 0.5022786458333331, "val_editor_changed_fraction": 0.796875, "val_editor_avg_edits": 1.7109375, "val_editor_in_candidate_set_fraction": 0.2421875, "val_editor_gate_precision": 0.4885844748858447, "val_editor_gate_recall": 0.7039473684210527, "val_editor_gate_f1": 0.5768194070080862, "val_editor_gate_positive_rate": 0.03491709183673469, "val_editor_gate_target_positive_rate": 0.02423469387755102, "val_editor_threshold": 0.7, "val_editor_oracle_gap_recovered": -0.3283582089552239 }, { "epoch": 14, "train_loss": 5.0008292575366795, "train_gate_loss": 0.39178889960385277, "train_value_loss": 4.609040347859263, "train_init_gate_loss": 1.0803341865539551e-07, "train_op_gate_loss": 0.004805860581074219, "train_arg_gate_loss": 0.38698293112065585, "train_init_value_loss": 2.920077775954269, "train_op_value_loss": 0.010811670364779981, "train_arg_value_loss": 1.6781509023858234, "train_target_changed_slots": 1.2890625, "train_target_changed_fraction": 0.02630739795918369, "train_target_oracle_fraction": 0.853515625, "train_target_gold_fallback_fraction": 0.146484375, "train_target_base_fallback_fraction": 0.0, "val_selected_threshold": 0.7, "val_n": 128.0, "val_oracle_found_fraction": 0.8515625, "val_avg_candidates": 1299.0, "val_avg_positive_candidates": 1.2890625, "val_base_executor_accuracy": 0.328125, "val_base_program_exact": 0.328125, "val_base_state_prefix_fraction": 0.5657552083333333, "val_base_changed_fraction": 0.0, "val_base_avg_edits": 0.0, "val_base_in_candidate_set_fraction": 1.0, "val_oracle_executor_accuracy": 0.8515625, "val_oracle_program_exact": 0.84375, "val_oracle_state_prefix_fraction": 0.8753255208333336, "val_oracle_changed_fraction": 0.5234375, "val_oracle_avg_edits": 0.6953125, "val_oracle_in_candidate_set_fraction": 1.0, "val_editor_executor_accuracy": 0.3046875, "val_editor_program_exact": 0.3046875, "val_editor_state_prefix_fraction": 0.5582682291666666, "val_editor_changed_fraction": 0.6484375, "val_editor_avg_edits": 1.265625, "val_editor_in_candidate_set_fraction": 0.4375, "val_editor_gate_precision": 0.5705521472392638, "val_editor_gate_recall": 0.6118421052631579, "val_editor_gate_f1": 0.5904761904761906, "val_editor_gate_positive_rate": 0.025988520408163265, "val_editor_gate_target_positive_rate": 0.02423469387755102, "val_editor_threshold": 0.7, "val_editor_oracle_gap_recovered": -0.04477611940298507 }, { "epoch": 15, "train_loss": 4.915371345123276, "train_gate_loss": 0.3928104439364688, "train_value_loss": 4.522560914978385, "train_init_gate_loss": 4.0978193283081055e-08, "train_op_gate_loss": 0.0043267152650423996, "train_arg_gate_loss": 0.38848368688559276, "train_init_value_loss": 2.857452351017855, "train_op_value_loss": 0.011145318850810826, "train_arg_value_loss": 1.6539632399799302, "train_target_changed_slots": 1.2890625, "train_target_changed_fraction": 0.026307397959183656, "train_target_oracle_fraction": 0.853515625, "train_target_gold_fallback_fraction": 0.146484375, "train_target_base_fallback_fraction": 0.0, "val_selected_threshold": 0.7, "val_n": 128.0, "val_oracle_found_fraction": 0.8515625, "val_avg_candidates": 1299.0, "val_avg_positive_candidates": 1.2890625, "val_base_executor_accuracy": 0.328125, "val_base_program_exact": 0.328125, "val_base_state_prefix_fraction": 0.5657552083333333, "val_base_changed_fraction": 0.0, "val_base_avg_edits": 0.0, "val_base_in_candidate_set_fraction": 1.0, "val_oracle_executor_accuracy": 0.8515625, "val_oracle_program_exact": 0.84375, "val_oracle_state_prefix_fraction": 0.8753255208333336, "val_oracle_changed_fraction": 0.5234375, "val_oracle_avg_edits": 0.6953125, "val_oracle_in_candidate_set_fraction": 1.0, "val_editor_executor_accuracy": 0.1953125, "val_editor_program_exact": 0.1953125, "val_editor_state_prefix_fraction": 0.4928385416666666, "val_editor_changed_fraction": 0.7890625, "val_editor_avg_edits": 1.78125, "val_editor_in_candidate_set_fraction": 0.2734375, "val_editor_gate_precision": 0.4517543859649123, "val_editor_gate_recall": 0.6776315789473685, "val_editor_gate_f1": 0.5421052631578948, "val_editor_gate_positive_rate": 0.03635204081632653, "val_editor_gate_target_positive_rate": 0.02423469387755102, "val_editor_threshold": 0.7, "val_editor_oracle_gap_recovered": -0.2537313432835821 }, { "epoch": 16, "train_loss": 4.776502750813961, "train_gate_loss": 0.3923414874006994, "train_value_loss": 4.3841612718533725, "train_init_gate_loss": 1.862645149230957e-09, "train_op_gate_loss": 0.004772133722445733, "train_arg_gate_loss": 0.3875693522131769, "train_init_value_loss": 2.73806291987421, "train_op_value_loss": 0.008663103718842945, "train_arg_value_loss": 1.6374352481216192, "train_target_changed_slots": 1.2890625, "train_target_changed_fraction": 0.026307397959183684, "train_target_oracle_fraction": 0.853515625, "train_target_gold_fallback_fraction": 0.146484375, "train_target_base_fallback_fraction": 0.0, "val_selected_threshold": 0.7, "val_n": 128.0, "val_oracle_found_fraction": 0.8515625, "val_avg_candidates": 1299.0, "val_avg_positive_candidates": 1.2890625, "val_base_executor_accuracy": 0.328125, "val_base_program_exact": 0.328125, "val_base_state_prefix_fraction": 0.5657552083333333, "val_base_changed_fraction": 0.0, "val_base_avg_edits": 0.0, "val_base_in_candidate_set_fraction": 1.0, "val_oracle_executor_accuracy": 0.8515625, "val_oracle_program_exact": 0.84375, "val_oracle_state_prefix_fraction": 0.8753255208333336, "val_oracle_changed_fraction": 0.5234375, "val_oracle_avg_edits": 0.6953125, "val_oracle_in_candidate_set_fraction": 1.0, "val_editor_executor_accuracy": 0.328125, "val_editor_program_exact": 0.328125, "val_editor_state_prefix_fraction": 0.6155598958333333, "val_editor_changed_fraction": 0.5859375, "val_editor_avg_edits": 1.0234375, "val_editor_in_candidate_set_fraction": 0.515625, "val_editor_gate_precision": 0.6870229007633588, "val_editor_gate_recall": 0.5921052631578947, "val_editor_gate_f1": 0.6360424028268551, "val_editor_gate_positive_rate": 0.020886479591836735, "val_editor_gate_target_positive_rate": 0.02423469387755102, "val_editor_threshold": 0.7, "val_editor_oracle_gap_recovered": 0.0 }, { "epoch": 17, "train_loss": 4.687056303257123, "train_gate_loss": 0.38643144942579966, "train_value_loss": 4.3006248557940125, "train_init_gate_loss": 1.862645149230957e-08, "train_op_gate_loss": 0.005019248298211848, "train_arg_gate_loss": 0.3814121826162591, "train_init_value_loss": 2.686028102529235, "train_op_value_loss": 0.013974834674129629, "train_arg_value_loss": 1.6006219265982509, "train_target_changed_slots": 1.2890625, "train_target_changed_fraction": 0.026307397959183694, "train_target_oracle_fraction": 0.853515625, "train_target_gold_fallback_fraction": 0.146484375, "train_target_base_fallback_fraction": 0.0, "val_selected_threshold": 0.7, "val_n": 128.0, "val_oracle_found_fraction": 0.8515625, "val_avg_candidates": 1299.0, "val_avg_positive_candidates": 1.2890625, "val_base_executor_accuracy": 0.328125, "val_base_program_exact": 0.328125, "val_base_state_prefix_fraction": 0.5657552083333333, "val_base_changed_fraction": 0.0, "val_base_avg_edits": 0.0, "val_base_in_candidate_set_fraction": 1.0, "val_oracle_executor_accuracy": 0.8515625, "val_oracle_program_exact": 0.84375, "val_oracle_state_prefix_fraction": 0.8753255208333336, "val_oracle_changed_fraction": 0.5234375, "val_oracle_avg_edits": 0.6953125, "val_oracle_in_candidate_set_fraction": 1.0, "val_editor_executor_accuracy": 0.2734375, "val_editor_program_exact": 0.265625, "val_editor_state_prefix_fraction": 0.5462239583333331, "val_editor_changed_fraction": 0.71875, "val_editor_avg_edits": 1.3984375, "val_editor_in_candidate_set_fraction": 0.375, "val_editor_gate_precision": 0.5611111111111111, "val_editor_gate_recall": 0.6644736842105263, "val_editor_gate_f1": 0.608433734939759, "val_editor_gate_positive_rate": 0.028698979591836735, "val_editor_gate_target_positive_rate": 0.02423469387755102, "val_editor_threshold": 0.7, "val_editor_oracle_gap_recovered": -0.1044776119402985 }, { "epoch": 18, "train_loss": 4.675375485327095, "train_gate_loss": 0.38388192880665883, "train_value_loss": 4.291493552038446, "train_init_gate_loss": 3.725290298461914e-09, "train_op_gate_loss": 0.004687460994680863, "train_arg_gate_loss": 0.37919446408341173, "train_init_value_loss": 2.6885953208548017, "train_op_value_loss": 0.01090801099328118, "train_arg_value_loss": 1.591990218614228, "train_target_changed_slots": 1.2890625, "train_target_changed_fraction": 0.02630739795918371, "train_target_oracle_fraction": 0.853515625, "train_target_gold_fallback_fraction": 0.146484375, "train_target_base_fallback_fraction": 0.0, "val_selected_threshold": 0.7, "val_n": 128.0, "val_oracle_found_fraction": 0.8515625, "val_avg_candidates": 1299.0, "val_avg_positive_candidates": 1.2890625, "val_base_executor_accuracy": 0.328125, "val_base_program_exact": 0.328125, "val_base_state_prefix_fraction": 0.5657552083333333, "val_base_changed_fraction": 0.0, "val_base_avg_edits": 0.0, "val_base_in_candidate_set_fraction": 1.0, "val_oracle_executor_accuracy": 0.8515625, "val_oracle_program_exact": 0.84375, "val_oracle_state_prefix_fraction": 0.8753255208333336, "val_oracle_changed_fraction": 0.5234375, "val_oracle_avg_edits": 0.6953125, "val_oracle_in_candidate_set_fraction": 1.0, "val_editor_executor_accuracy": 0.3515625, "val_editor_program_exact": 0.3515625, "val_editor_state_prefix_fraction": 0.6201171874999999, "val_editor_changed_fraction": 0.6640625, "val_editor_avg_edits": 1.1640625, "val_editor_in_candidate_set_fraction": 0.484375, "val_editor_gate_precision": 0.66, "val_editor_gate_recall": 0.6513157894736842, "val_editor_gate_f1": 0.6556291390728477, "val_editor_gate_positive_rate": 0.02391581632653061, "val_editor_gate_target_positive_rate": 0.02423469387755102, "val_editor_threshold": 0.7, "val_editor_oracle_gap_recovered": 0.04477611940298507 } ] }