{
  "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": "pilot_slot_repair_gated_s96",
      "compiler_checkpoint": "large_artifacts/qwen_slot_repair_distillation/checkpoints/fixed_compiler_step00800",
      "train_examples": 96,
      "val_examples": 48,
      "eval_examples": 64,
      "eval_pairs": 64,
      "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": 96,
      "editor_layers": 2,
      "editor_heads": 4,
      "editor_ff_mult": 4,
      "editor_dropout": 0.05,
      "editor_epochs": 10,
      "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": 0.0,
      "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": 63,
      "eval_seed": 63001
    },
    "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": 3,
    "best_editor_threshold": 0.6,
    "train_seconds": 183.1253514289856
  },
  "metrics": {
    "train_len24": {
      "n": 96.0,
      "oracle_found_fraction": 0.875,
      "avg_candidates": 1299.0,
      "avg_positive_candidates": 1.2708333333333333,
      "base_executor_accuracy": 0.2708333333333333,
      "base_program_exact": 0.2708333333333333,
      "base_state_prefix_fraction": 0.5685763888888887,
      "base_changed_fraction": 0.0,
      "base_avg_edits": 0.0,
      "base_in_candidate_set_fraction": 1.0,
      "oracle_executor_accuracy": 0.875,
      "oracle_program_exact": 0.875,
      "oracle_state_prefix_fraction": 0.8993055555555557,
      "oracle_changed_fraction": 0.6041666666666666,
      "oracle_avg_edits": 0.8125,
      "oracle_in_candidate_set_fraction": 1.0,
      "editor_executor_accuracy": 0.2916666666666667,
      "editor_program_exact": 0.2916666666666667,
      "editor_state_prefix_fraction": 0.5850694444444443,
      "editor_changed_fraction": 0.22916666666666666,
      "editor_avg_edits": 0.2708333333333333,
      "editor_in_candidate_set_fraction": 0.8125,
      "editor_gate_precision": 0.8846153846153846,
      "editor_gate_recall": 0.19327731092436976,
      "editor_gate_f1": 0.31724137931034485,
      "editor_gate_positive_rate": 0.0055272108843537416,
      "editor_gate_target_positive_rate": 0.025297619047619048,
      "editor_threshold": 0.6,
      "editor_oracle_gap_recovered": 0.03448275862068971
    },
    "val_len24": {
      "n": 48.0,
      "oracle_found_fraction": 0.8541666666666666,
      "avg_candidates": 1299.0,
      "avg_positive_candidates": 1.1458333333333333,
      "base_executor_accuracy": 0.25,
      "base_program_exact": 0.25,
      "base_state_prefix_fraction": 0.537326388888889,
      "base_changed_fraction": 0.0,
      "base_avg_edits": 0.0,
      "base_in_candidate_set_fraction": 1.0,
      "oracle_executor_accuracy": 0.8541666666666666,
      "oracle_program_exact": 0.8541666666666666,
      "oracle_state_prefix_fraction": 0.8897569444444445,
      "oracle_changed_fraction": 0.6041666666666666,
      "oracle_avg_edits": 0.8541666666666666,
      "oracle_in_candidate_set_fraction": 1.0,
      "editor_executor_accuracy": 0.3125,
      "editor_program_exact": 0.3125,
      "editor_state_prefix_fraction": 0.5798611111111113,
      "editor_changed_fraction": 0.14583333333333334,
      "editor_avg_edits": 0.14583333333333334,
      "editor_in_candidate_set_fraction": 0.9166666666666666,
      "editor_gate_precision": 0.8571428571428571,
      "editor_gate_recall": 0.0967741935483871,
      "editor_gate_f1": 0.17391304347826084,
      "editor_gate_positive_rate": 0.002976190476190476,
      "editor_gate_target_positive_rate": 0.026360544217687076,
      "editor_threshold": 0.6,
      "editor_oracle_gap_recovered": 0.10344827586206898
    },
    "fresh_standard_len24": {
      "n": 64.0,
      "oracle_found_fraction": 0.90625,
      "avg_candidates": 1299.0,
      "avg_positive_candidates": 1.375,
      "base_executor_accuracy": 0.359375,
      "base_program_exact": 0.359375,
      "base_state_prefix_fraction": 0.5911458333333335,
      "base_changed_fraction": 0.0,
      "base_avg_edits": 0.0,
      "base_in_candidate_set_fraction": 1.0,
      "oracle_executor_accuracy": 0.90625,
      "oracle_program_exact": 0.90625,
      "oracle_state_prefix_fraction": 0.9270833333333334,
      "oracle_changed_fraction": 0.546875,
      "oracle_avg_edits": 0.828125,
      "oracle_in_candidate_set_fraction": 1.0,
      "editor_executor_accuracy": 0.34375,
      "editor_program_exact": 0.34375,
      "editor_state_prefix_fraction": 0.587890625,
      "editor_changed_fraction": 0.109375,
      "editor_avg_edits": 0.125,
      "editor_in_candidate_set_fraction": 0.90625,
      "editor_gate_precision": 0.875,
      "editor_gate_recall": 0.0958904109589041,
      "editor_gate_f1": 0.17283950617283952,
      "editor_gate_positive_rate": 0.002551020408163265,
      "editor_gate_target_positive_rate": 0.023278061224489797,
      "editor_threshold": 0.6,
      "editor_oracle_gap_recovered": -0.02857142857142857
    },
    "fresh_paraphrase_len24": {
      "n": 64.0,
      "oracle_found_fraction": 0.890625,
      "avg_candidates": 1299.0,
      "avg_positive_candidates": 1.28125,
      "base_executor_accuracy": 0.1875,
      "base_program_exact": 0.1875,
      "base_state_prefix_fraction": 0.5143229166666666,
      "base_changed_fraction": 0.0,
      "base_avg_edits": 0.0,
      "base_in_candidate_set_fraction": 1.0,
      "oracle_executor_accuracy": 0.890625,
      "oracle_program_exact": 0.890625,
      "oracle_state_prefix_fraction": 0.9114583333333334,
      "oracle_changed_fraction": 0.703125,
      "oracle_avg_edits": 0.953125,
      "oracle_in_candidate_set_fraction": 1.0,
      "editor_executor_accuracy": 0.171875,
      "editor_program_exact": 0.171875,
      "editor_state_prefix_fraction": 0.5110677083333333,
      "editor_changed_fraction": 0.09375,
      "editor_avg_edits": 0.109375,
      "editor_in_candidate_set_fraction": 0.921875,
      "editor_gate_precision": 0.5714285714285714,
      "editor_gate_recall": 0.047619047619047616,
      "editor_gate_f1": 0.08791208791208789,
      "editor_gate_positive_rate": 0.002232142857142857,
      "editor_gate_target_positive_rate": 0.026785714285714284,
      "editor_threshold": 0.6,
      "editor_oracle_gap_recovered": -0.022222222222222223
    },
    "fresh_paired_len24": {
      "n": 128.0,
      "oracle_found_fraction": 0.8671875,
      "avg_candidates": 1299.0,
      "avg_positive_candidates": 2.1015625,
      "base_executor_accuracy": 0.3046875,
      "base_program_exact": 0.3046875,
      "base_state_prefix_fraction": 0.5777994791666665,
      "base_changed_fraction": 0.0,
      "base_avg_edits": 0.0,
      "base_in_candidate_set_fraction": 1.0,
      "oracle_executor_accuracy": 0.8671875,
      "oracle_program_exact": 0.8671875,
      "oracle_state_prefix_fraction": 0.8909505208333333,
      "oracle_changed_fraction": 0.5625,
      "oracle_avg_edits": 0.8046875,
      "oracle_in_candidate_set_fraction": 1.0,
      "editor_executor_accuracy": 0.28125,
      "editor_program_exact": 0.28125,
      "editor_state_prefix_fraction": 0.5690104166666665,
      "editor_changed_fraction": 0.140625,
      "editor_avg_edits": 0.1796875,
      "editor_in_candidate_set_fraction": 0.890625,
      "editor_gate_precision": 0.782608695652174,
      "editor_gate_recall": 0.10909090909090909,
      "editor_gate_f1": 0.19148936170212763,
      "editor_gate_positive_rate": 0.003667091836734694,
      "editor_gate_target_positive_rate": 0.026307397959183673,
      "editor_threshold": 0.6,
      "editor_oracle_gap_recovered": -0.041666666666666664,
      "base_pair_answer_consistency": 0.703125,
      "base_pair_both_correct": 0.28125,
      "base_pair_program_consistency": 0.703125,
      "base_pair_state_consistency": 0.703125,
      "oracle_pair_answer_consistency": 0.875,
      "oracle_pair_both_correct": 0.84375,
      "oracle_pair_program_consistency": 0.875,
      "oracle_pair_state_consistency": 0.875,
      "editor_pair_answer_consistency": 0.59375,
      "editor_pair_both_correct": 0.234375,
      "editor_pair_program_consistency": 0.59375,
      "editor_pair_state_consistency": 0.59375
    }
  },
  "train_log": [
    {
      "epoch": 1,
      "train_loss": 2.6394869547802955,
      "train_gate_loss": 0.9412104461807758,
      "train_value_loss": 1.698276499286294,
      "train_init_gate_loss": 0.007466276486714681,
      "train_op_gate_loss": 0.038686605086695636,
      "train_arg_gate_loss": 0.8950575699370044,
      "train_init_value_loss": 0.0,
      "train_op_value_loss": 0.013054573287566503,
      "train_arg_value_loss": 1.685221924756964,
      "train_target_changed_slots": 1.2395833333333333,
      "train_target_changed_fraction": 0.02529761904761904,
      "train_target_oracle_fraction": 0.875,
      "train_target_gold_fallback_fraction": 0.125,
      "train_target_base_fallback_fraction": 0.0,
      "val_selected_threshold": 0.25,
      "val_n": 48.0,
      "val_oracle_found_fraction": 0.8541666666666666,
      "val_avg_candidates": 1299.0,
      "val_avg_positive_candidates": 1.1458333333333333,
      "val_base_executor_accuracy": 0.25,
      "val_base_program_exact": 0.25,
      "val_base_state_prefix_fraction": 0.537326388888889,
      "val_base_changed_fraction": 0.0,
      "val_base_avg_edits": 0.0,
      "val_base_in_candidate_set_fraction": 1.0,
      "val_oracle_executor_accuracy": 0.8541666666666666,
      "val_oracle_program_exact": 0.8541666666666666,
      "val_oracle_state_prefix_fraction": 0.8897569444444445,
      "val_oracle_changed_fraction": 0.6041666666666666,
      "val_oracle_avg_edits": 0.8541666666666666,
      "val_oracle_in_candidate_set_fraction": 1.0,
      "val_editor_executor_accuracy": 0.2708333333333333,
      "val_editor_program_exact": 0.2708333333333333,
      "val_editor_state_prefix_fraction": 0.5460069444444445,
      "val_editor_changed_fraction": 0.020833333333333332,
      "val_editor_avg_edits": 0.020833333333333332,
      "val_editor_in_candidate_set_fraction": 1.0,
      "val_editor_gate_precision": 1.0,
      "val_editor_gate_recall": 0.016129032258064516,
      "val_editor_gate_f1": 0.031746031746031744,
      "val_editor_gate_positive_rate": 0.00042517006802721087,
      "val_editor_gate_target_positive_rate": 0.026360544217687076,
      "val_editor_threshold": 0.25,
      "val_editor_oracle_gap_recovered": 0.03448275862068963
    },
    {
      "epoch": 2,
      "train_loss": 1.8628560314537026,
      "train_gate_loss": 0.7490735122895179,
      "train_value_loss": 1.1137825213372707,
      "train_init_gate_loss": 0.00019637743631998697,
      "train_op_gate_loss": 0.029328587299157032,
      "train_arg_gate_loss": 0.7195485510164872,
      "train_init_value_loss": 0.0,
      "train_op_value_loss": 0.0027426229789853096,
      "train_arg_value_loss": 1.1110398980478446,
      "train_target_changed_slots": 1.2395833333333333,
      "train_target_changed_fraction": 0.02529761904761905,
      "train_target_oracle_fraction": 0.875,
      "train_target_gold_fallback_fraction": 0.125,
      "train_target_base_fallback_fraction": 0.0,
      "val_selected_threshold": 0.2,
      "val_n": 48.0,
      "val_oracle_found_fraction": 0.8541666666666666,
      "val_avg_candidates": 1299.0,
      "val_avg_positive_candidates": 1.1458333333333333,
      "val_base_executor_accuracy": 0.25,
      "val_base_program_exact": 0.25,
      "val_base_state_prefix_fraction": 0.537326388888889,
      "val_base_changed_fraction": 0.0,
      "val_base_avg_edits": 0.0,
      "val_base_in_candidate_set_fraction": 1.0,
      "val_oracle_executor_accuracy": 0.8541666666666666,
      "val_oracle_program_exact": 0.8541666666666666,
      "val_oracle_state_prefix_fraction": 0.8897569444444445,
      "val_oracle_changed_fraction": 0.6041666666666666,
      "val_oracle_avg_edits": 0.8541666666666666,
      "val_oracle_in_candidate_set_fraction": 1.0,
      "val_editor_executor_accuracy": 0.25,
      "val_editor_program_exact": 0.25,
      "val_editor_state_prefix_fraction": 0.5434027777777779,
      "val_editor_changed_fraction": 0.4791666666666667,
      "val_editor_avg_edits": 0.7083333333333334,
      "val_editor_in_candidate_set_fraction": 0.5833333333333334,
      "val_editor_gate_precision": 0.5588235294117647,
      "val_editor_gate_recall": 0.3064516129032258,
      "val_editor_gate_f1": 0.39583333333333337,
      "val_editor_gate_positive_rate": 0.01445578231292517,
      "val_editor_gate_target_positive_rate": 0.026360544217687076,
      "val_editor_threshold": 0.2,
      "val_editor_oracle_gap_recovered": 0.0
    },
    {
      "epoch": 3,
      "train_loss": 1.6753169826697558,
      "train_gate_loss": 0.6827187992166728,
      "train_value_loss": 0.9925981843844056,
      "train_init_gate_loss": 8.046130339304607e-05,
      "train_op_gate_loss": 0.0279772744473424,
      "train_arg_gate_loss": 0.6546610625033887,
      "train_init_value_loss": 0.0,
      "train_op_value_loss": 0.0007831741434832414,
      "train_arg_value_loss": 0.9918150091543794,
      "train_target_changed_slots": 1.2395833333333333,
      "train_target_changed_fraction": 0.025297619047619055,
      "train_target_oracle_fraction": 0.875,
      "train_target_gold_fallback_fraction": 0.125,
      "train_target_base_fallback_fraction": 0.0,
      "val_selected_threshold": 0.6,
      "val_n": 48.0,
      "val_oracle_found_fraction": 0.8541666666666666,
      "val_avg_candidates": 1299.0,
      "val_avg_positive_candidates": 1.1458333333333333,
      "val_base_executor_accuracy": 0.25,
      "val_base_program_exact": 0.25,
      "val_base_state_prefix_fraction": 0.537326388888889,
      "val_base_changed_fraction": 0.0,
      "val_base_avg_edits": 0.0,
      "val_base_in_candidate_set_fraction": 1.0,
      "val_oracle_executor_accuracy": 0.8541666666666666,
      "val_oracle_program_exact": 0.8541666666666666,
      "val_oracle_state_prefix_fraction": 0.8897569444444445,
      "val_oracle_changed_fraction": 0.6041666666666666,
      "val_oracle_avg_edits": 0.8541666666666666,
      "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": 0.5798611111111113,
      "val_editor_changed_fraction": 0.14583333333333334,
      "val_editor_avg_edits": 0.14583333333333334,
      "val_editor_in_candidate_set_fraction": 0.9166666666666666,
      "val_editor_gate_precision": 0.8571428571428571,
      "val_editor_gate_recall": 0.0967741935483871,
      "val_editor_gate_f1": 0.17391304347826084,
      "val_editor_gate_positive_rate": 0.002976190476190476,
      "val_editor_gate_target_positive_rate": 0.026360544217687076,
      "val_editor_threshold": 0.6,
      "val_editor_oracle_gap_recovered": 0.10344827586206898
    },
    {
      "epoch": 4,
      "train_loss": 1.5090648010203342,
      "train_gate_loss": 0.5424656952600344,
      "train_value_loss": 0.9665991120661298,
      "train_init_gate_loss": 3.808736801147461e-05,
      "train_op_gate_loss": 0.02130127591438698,
      "train_arg_gate_loss": 0.5211263304324044,
      "train_init_value_loss": 0.0,
      "train_op_value_loss": 0.00034430588129907846,
      "train_arg_value_loss": 0.9662548058355848,
      "train_target_changed_slots": 1.2395833333333333,
      "train_target_changed_fraction": 0.025297619047619038,
      "train_target_oracle_fraction": 0.875,
      "train_target_gold_fallback_fraction": 0.125,
      "train_target_base_fallback_fraction": 0.0,
      "val_selected_threshold": 0.7,
      "val_n": 48.0,
      "val_oracle_found_fraction": 0.8541666666666666,
      "val_avg_candidates": 1299.0,
      "val_avg_positive_candidates": 1.1458333333333333,
      "val_base_executor_accuracy": 0.25,
      "val_base_program_exact": 0.25,
      "val_base_state_prefix_fraction": 0.537326388888889,
      "val_base_changed_fraction": 0.0,
      "val_base_avg_edits": 0.0,
      "val_base_in_candidate_set_fraction": 1.0,
      "val_oracle_executor_accuracy": 0.8541666666666666,
      "val_oracle_program_exact": 0.8541666666666666,
      "val_oracle_state_prefix_fraction": 0.8897569444444445,
      "val_oracle_changed_fraction": 0.6041666666666666,
      "val_oracle_avg_edits": 0.8541666666666666,
      "val_oracle_in_candidate_set_fraction": 1.0,
      "val_editor_executor_accuracy": 0.2708333333333333,
      "val_editor_program_exact": 0.2708333333333333,
      "val_editor_state_prefix_fraction": 0.5451388888888891,
      "val_editor_changed_fraction": 0.125,
      "val_editor_avg_edits": 0.125,
      "val_editor_in_candidate_set_fraction": 0.9166666666666666,
      "val_editor_gate_precision": 0.6666666666666666,
      "val_editor_gate_recall": 0.06451612903225806,
      "val_editor_gate_f1": 0.1176470588235294,
      "val_editor_gate_positive_rate": 0.002551020408163265,
      "val_editor_gate_target_positive_rate": 0.026360544217687076,
      "val_editor_threshold": 0.7,
      "val_editor_oracle_gap_recovered": 0.03448275862068963
    },
    {
      "epoch": 5,
      "train_loss": 1.473886678017152,
      "train_gate_loss": 0.49434317918348825,
      "train_value_loss": 0.979543499648571,
      "train_init_gate_loss": 3.49879264831543e-05,
      "train_op_gate_loss": 0.017772549367843265,
      "train_arg_gate_loss": 0.4765356480153666,
      "train_init_value_loss": 0.0,
      "train_op_value_loss": 0.00013881649162309864,
      "train_arg_value_loss": 0.979404682914416,
      "train_target_changed_slots": 1.2395833333333333,
      "train_target_changed_fraction": 0.02529761904761905,
      "train_target_oracle_fraction": 0.875,
      "train_target_gold_fallback_fraction": 0.125,
      "train_target_base_fallback_fraction": 0.0,
      "val_selected_threshold": 0.7,
      "val_n": 48.0,
      "val_oracle_found_fraction": 0.8541666666666666,
      "val_avg_candidates": 1299.0,
      "val_avg_positive_candidates": 1.1458333333333333,
      "val_base_executor_accuracy": 0.25,
      "val_base_program_exact": 0.25,
      "val_base_state_prefix_fraction": 0.537326388888889,
      "val_base_changed_fraction": 0.0,
      "val_base_avg_edits": 0.0,
      "val_base_in_candidate_set_fraction": 1.0,
      "val_oracle_executor_accuracy": 0.8541666666666666,
      "val_oracle_program_exact": 0.8541666666666666,
      "val_oracle_state_prefix_fraction": 0.8897569444444445,
      "val_oracle_changed_fraction": 0.6041666666666666,
      "val_oracle_avg_edits": 0.8541666666666666,
      "val_oracle_in_candidate_set_fraction": 1.0,
      "val_editor_executor_accuracy": 0.20833333333333334,
      "val_editor_program_exact": 0.20833333333333334,
      "val_editor_state_prefix_fraction": 0.5269097222222224,
      "val_editor_changed_fraction": 0.4583333333333333,
      "val_editor_avg_edits": 0.8333333333333334,
      "val_editor_in_candidate_set_fraction": 0.5625,
      "val_editor_gate_precision": 0.5,
      "val_editor_gate_recall": 0.3225806451612903,
      "val_editor_gate_f1": 0.39215686274509803,
      "val_editor_gate_positive_rate": 0.017006802721088437,
      "val_editor_gate_target_positive_rate": 0.026360544217687076,
      "val_editor_threshold": 0.7,
      "val_editor_oracle_gap_recovered": -0.0689655172413793
    },
    {
      "epoch": 6,
      "train_loss": 1.454232336156262,
      "train_gate_loss": 0.5622106231151823,
      "train_value_loss": 0.8920217140888175,
      "train_init_gate_loss": 4.372497399648031e-05,
      "train_op_gate_loss": 0.015886565643957812,
      "train_arg_gate_loss": 0.5462803320569947,
      "train_init_value_loss": 0.0,
      "train_op_value_loss": 7.420362574824442e-05,
      "train_arg_value_loss": 0.8919475112731258,
      "train_target_changed_slots": 1.2395833333333333,
      "train_target_changed_fraction": 0.025297619047619044,
      "train_target_oracle_fraction": 0.875,
      "train_target_gold_fallback_fraction": 0.125,
      "train_target_base_fallback_fraction": 0.0,
      "val_selected_threshold": 0.7,
      "val_n": 48.0,
      "val_oracle_found_fraction": 0.8541666666666666,
      "val_avg_candidates": 1299.0,
      "val_avg_positive_candidates": 1.1458333333333333,
      "val_base_executor_accuracy": 0.25,
      "val_base_program_exact": 0.25,
      "val_base_state_prefix_fraction": 0.537326388888889,
      "val_base_changed_fraction": 0.0,
      "val_base_avg_edits": 0.0,
      "val_base_in_candidate_set_fraction": 1.0,
      "val_oracle_executor_accuracy": 0.8541666666666666,
      "val_oracle_program_exact": 0.8541666666666666,
      "val_oracle_state_prefix_fraction": 0.8897569444444445,
      "val_oracle_changed_fraction": 0.6041666666666666,
      "val_oracle_avg_edits": 0.8541666666666666,
      "val_oracle_in_candidate_set_fraction": 1.0,
      "val_editor_executor_accuracy": 0.22916666666666666,
      "val_editor_program_exact": 0.22916666666666666,
      "val_editor_state_prefix_fraction": 0.5182291666666667,
      "val_editor_changed_fraction": 0.5625,
      "val_editor_avg_edits": 0.8333333333333334,
      "val_editor_in_candidate_set_fraction": 0.4791666666666667,
      "val_editor_gate_precision": 0.6,
      "val_editor_gate_recall": 0.3870967741935484,
      "val_editor_gate_f1": 0.47058823529411764,
      "val_editor_gate_positive_rate": 0.017006802721088437,
      "val_editor_gate_target_positive_rate": 0.026360544217687076,
      "val_editor_threshold": 0.7,
      "val_editor_oracle_gap_recovered": -0.034482758620689676
    },
    {
      "epoch": 7,
      "train_loss": 1.2398894230573205,
      "train_gate_loss": 0.40541489191915997,
      "train_value_loss": 0.8344745355813453,
      "train_init_gate_loss": 3.09149424235026e-05,
      "train_op_gate_loss": 0.013655037383273338,
      "train_arg_gate_loss": 0.3917289434418005,
      "train_init_value_loss": 0.0,
      "train_op_value_loss": 0.00014460331294685602,
      "train_arg_value_loss": 0.8343299322295934,
      "train_target_changed_slots": 1.2395833333333333,
      "train_target_changed_fraction": 0.025297619047619044,
      "train_target_oracle_fraction": 0.875,
      "train_target_gold_fallback_fraction": 0.125,
      "train_target_base_fallback_fraction": 0.0,
      "val_selected_threshold": 0.7,
      "val_n": 48.0,
      "val_oracle_found_fraction": 0.8541666666666666,
      "val_avg_candidates": 1299.0,
      "val_avg_positive_candidates": 1.1458333333333333,
      "val_base_executor_accuracy": 0.25,
      "val_base_program_exact": 0.25,
      "val_base_state_prefix_fraction": 0.537326388888889,
      "val_base_changed_fraction": 0.0,
      "val_base_avg_edits": 0.0,
      "val_base_in_candidate_set_fraction": 1.0,
      "val_oracle_executor_accuracy": 0.8541666666666666,
      "val_oracle_program_exact": 0.8541666666666666,
      "val_oracle_state_prefix_fraction": 0.8897569444444445,
      "val_oracle_changed_fraction": 0.6041666666666666,
      "val_oracle_avg_edits": 0.8541666666666666,
      "val_oracle_in_candidate_set_fraction": 1.0,
      "val_editor_executor_accuracy": 0.2708333333333333,
      "val_editor_program_exact": 0.2708333333333333,
      "val_editor_state_prefix_fraction": 0.5251736111111112,
      "val_editor_changed_fraction": 0.4791666666666667,
      "val_editor_avg_edits": 0.8541666666666666,
      "val_editor_in_candidate_set_fraction": 0.5416666666666666,
      "val_editor_gate_precision": 0.4883720930232558,
      "val_editor_gate_recall": 0.3387096774193548,
      "val_editor_gate_f1": 0.4,
      "val_editor_gate_positive_rate": 0.018282312925170068,
      "val_editor_gate_target_positive_rate": 0.026360544217687076,
      "val_editor_threshold": 0.7,
      "val_editor_oracle_gap_recovered": 0.03448275862068963
    },
    {
      "epoch": 8,
      "train_loss": 1.2704105892016742,
      "train_gate_loss": 0.41250028598703165,
      "train_value_loss": 0.857910306736206,
      "train_init_gate_loss": 2.919634183247884e-05,
      "train_op_gate_loss": 0.008917635548674715,
      "train_arg_gate_loss": 0.4035534521341712,
      "train_init_value_loss": 0.0,
      "train_op_value_loss": 4.668858309742063e-05,
      "train_arg_value_loss": 0.8578636189146588,
      "train_target_changed_slots": 1.2395833333333333,
      "train_target_changed_fraction": 0.02529761904761904,
      "train_target_oracle_fraction": 0.875,
      "train_target_gold_fallback_fraction": 0.125,
      "train_target_base_fallback_fraction": 0.0,
      "val_selected_threshold": 0.7,
      "val_n": 48.0,
      "val_oracle_found_fraction": 0.8541666666666666,
      "val_avg_candidates": 1299.0,
      "val_avg_positive_candidates": 1.1458333333333333,
      "val_base_executor_accuracy": 0.25,
      "val_base_program_exact": 0.25,
      "val_base_state_prefix_fraction": 0.537326388888889,
      "val_base_changed_fraction": 0.0,
      "val_base_avg_edits": 0.0,
      "val_base_in_candidate_set_fraction": 1.0,
      "val_oracle_executor_accuracy": 0.8541666666666666,
      "val_oracle_program_exact": 0.8541666666666666,
      "val_oracle_state_prefix_fraction": 0.8897569444444445,
      "val_oracle_changed_fraction": 0.6041666666666666,
      "val_oracle_avg_edits": 0.8541666666666666,
      "val_oracle_in_candidate_set_fraction": 1.0,
      "val_editor_executor_accuracy": 0.25,
      "val_editor_program_exact": 0.25,
      "val_editor_state_prefix_fraction": 0.5642361111111113,
      "val_editor_changed_fraction": 0.4583333333333333,
      "val_editor_avg_edits": 0.6041666666666666,
      "val_editor_in_candidate_set_fraction": 0.6666666666666666,
      "val_editor_gate_precision": 0.7586206896551724,
      "val_editor_gate_recall": 0.3548387096774194,
      "val_editor_gate_f1": 0.4835164835164835,
      "val_editor_gate_positive_rate": 0.012329931972789115,
      "val_editor_gate_target_positive_rate": 0.026360544217687076,
      "val_editor_threshold": 0.7,
      "val_editor_oracle_gap_recovered": 0.0
    },
    {
      "epoch": 9,
      "train_loss": 1.1146585046617474,
      "train_gate_loss": 0.37367194377808727,
      "train_value_loss": 0.7409865644294769,
      "train_init_gate_loss": 4.0625532468159996e-05,
      "train_op_gate_loss": 0.009301952512027848,
      "train_arg_gate_loss": 0.36432936833989515,
      "train_init_value_loss": 0.0,
      "train_op_value_loss": 2.2828382498119026e-05,
      "train_arg_value_loss": 0.7409637358505279,
      "train_target_changed_slots": 1.2395833333333333,
      "train_target_changed_fraction": 0.02529761904761904,
      "train_target_oracle_fraction": 0.875,
      "train_target_gold_fallback_fraction": 0.125,
      "train_target_base_fallback_fraction": 0.0,
      "val_selected_threshold": 0.7,
      "val_n": 48.0,
      "val_oracle_found_fraction": 0.8541666666666666,
      "val_avg_candidates": 1299.0,
      "val_avg_positive_candidates": 1.1458333333333333,
      "val_base_executor_accuracy": 0.25,
      "val_base_program_exact": 0.25,
      "val_base_state_prefix_fraction": 0.537326388888889,
      "val_base_changed_fraction": 0.0,
      "val_base_avg_edits": 0.0,
      "val_base_in_candidate_set_fraction": 1.0,
      "val_oracle_executor_accuracy": 0.8541666666666666,
      "val_oracle_program_exact": 0.8541666666666666,
      "val_oracle_state_prefix_fraction": 0.8897569444444445,
      "val_oracle_changed_fraction": 0.6041666666666666,
      "val_oracle_avg_edits": 0.8541666666666666,
      "val_oracle_in_candidate_set_fraction": 1.0,
      "val_editor_executor_accuracy": 0.16666666666666666,
      "val_editor_program_exact": 0.16666666666666666,
      "val_editor_state_prefix_fraction": 0.5425347222222222,
      "val_editor_changed_fraction": 0.6458333333333334,
      "val_editor_avg_edits": 1.1041666666666667,
      "val_editor_in_candidate_set_fraction": 0.4166666666666667,
      "val_editor_gate_precision": 0.45454545454545453,
      "val_editor_gate_recall": 0.4032258064516129,
      "val_editor_gate_f1": 0.42735042735042733,
      "val_editor_gate_positive_rate": 0.0233843537414966,
      "val_editor_gate_target_positive_rate": 0.026360544217687076,
      "val_editor_threshold": 0.7,
      "val_editor_oracle_gap_recovered": -0.13793103448275865
    },
    {
      "epoch": 10,
      "train_loss": 1.186833067451668,
      "train_gate_loss": 0.4079064455037648,
      "train_value_loss": 0.7789266204248028,
      "train_init_gate_loss": 1.2576580047607422e-05,
      "train_op_gate_loss": 0.007386289553783361,
      "train_arg_gate_loss": 0.4005075783409969,
      "train_init_value_loss": 0.0,
      "train_op_value_loss": 1.802969442602868e-05,
      "train_arg_value_loss": 0.7789085912615216,
      "train_target_changed_slots": 1.2395833333333333,
      "train_target_changed_fraction": 0.025297619047619044,
      "train_target_oracle_fraction": 0.875,
      "train_target_gold_fallback_fraction": 0.125,
      "train_target_base_fallback_fraction": 0.0,
      "val_selected_threshold": 0.7,
      "val_n": 48.0,
      "val_oracle_found_fraction": 0.8541666666666666,
      "val_avg_candidates": 1299.0,
      "val_avg_positive_candidates": 1.1458333333333333,
      "val_base_executor_accuracy": 0.25,
      "val_base_program_exact": 0.25,
      "val_base_state_prefix_fraction": 0.537326388888889,
      "val_base_changed_fraction": 0.0,
      "val_base_avg_edits": 0.0,
      "val_base_in_candidate_set_fraction": 1.0,
      "val_oracle_executor_accuracy": 0.8541666666666666,
      "val_oracle_program_exact": 0.8541666666666666,
      "val_oracle_state_prefix_fraction": 0.8897569444444445,
      "val_oracle_changed_fraction": 0.6041666666666666,
      "val_oracle_avg_edits": 0.8541666666666666,
      "val_oracle_in_candidate_set_fraction": 1.0,
      "val_editor_executor_accuracy": 0.22916666666666666,
      "val_editor_program_exact": 0.22916666666666666,
      "val_editor_state_prefix_fraction": 0.5156250000000001,
      "val_editor_changed_fraction": 0.3541666666666667,
      "val_editor_avg_edits": 0.5625,
      "val_editor_in_candidate_set_fraction": 0.6666666666666666,
      "val_editor_gate_precision": 0.6666666666666666,
      "val_editor_gate_recall": 0.2903225806451613,
      "val_editor_gate_f1": 0.4044943820224719,
      "val_editor_gate_positive_rate": 0.011479591836734694,
      "val_editor_gate_target_positive_rate": 0.026360544217687076,
      "val_editor_threshold": 0.7,
      "val_editor_oracle_gap_recovered": -0.034482758620689676
    }
  ]
}