{
  "checkpoint": {
    "adapter_path": "large_artifacts/qwen_hidden_vm_curriculum_repair/checkpoints/smoke_curriculum_repair/adapter",
    "heads_path": "large_artifacts/qwen_hidden_vm_curriculum_repair/checkpoints/smoke_curriculum_repair/heads.pt",
    "run": "smoke_curriculum_repair"
  },
  "metadata": {
    "args": {
      "arg_distance_temperature": 4.0,
      "arg_selection_loss_weight": 4.0,
      "arg_trace_loss_weight": 4.0,
      "arg_window": 16,
      "curriculum_schedule": "2:1,3:2",
      "device_map": "auto",
      "direct_answer_loss_weight": 0.0,
      "domain_eval_examples": 2,
      "domains": "arithmetic,calendar,unit,list,boolean,lookup",
      "eval_batch_size": 2,
      "eval_every": 0,
      "eval_examples": 6,
      "eval_length": 3,
      "eval_pairs": 4,
      "eval_seed": 101001,
      "executor_loss_weight": 0.2,
      "grad_clip": 1.0,
      "gradient_checkpointing": true,
      "hard_length": 4,
      "harder_length": 0,
      "head_width": 512,
      "init_selection_loss_weight": 1.0,
      "init_trace_loss_weight": 4.0,
      "load_in_4bit": true,
      "log_every": 1,
      "lora_alpha": 16,
      "lora_dropout": 0.05,
      "lora_r": 8,
      "lora_target_modules": "all-linear",
      "lr": 5e-05,
      "max_length": 384,
      "max_steps": 4,
      "model_id": "Qwen/Qwen3-4B",
      "op_selection_loss_weight": 1.0,
      "op_trace_loss_weight": 1.0,
      "rank_temperature": 1.0,
      "repair_eval_topk": 2,
      "repair_include_unchanged": false,
      "repair_max_edits": 1,
      "repair_selection_loss_weight": 0.0,
      "repair_source_examples": 12,
      "repair_steps": 1,
      "repair_target_mode": "repair_only",
      "repair_train_topk": 2,
      "role_pos_weight": 20.0,
      "run_name": "smoke_curriculum_repair",
      "seed": 101,
      "selection_loss_weight": 1.0,
      "state_loss_weight": 0.05,
      "torch_dtype": "bf16",
      "trace_loss_weight": 1.0,
      "train_batch_size": 1,
      "train_examples": 12,
      "train_lora": true,
      "train_max_len": 3,
      "train_min_len": 1,
      "train_steps": 2,
      "train_template_mode": "mixed",
      "val_examples": 6,
      "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": 16.83807611465454
  },
  "metrics": {
    "domain_arithmetic": {
      "arg_accuracy": 0.6666666666666666,
      "direct_accuracy": 0.0,
      "domain_arithmetic_direct_accuracy": 0.0,
      "domain_arithmetic_executor_accuracy": 0.5,
      "domain_arithmetic_n": 2.0,
      "domain_arithmetic_repair_executor_accuracy": 0.5,
      "executor_accuracy": 0.5,
      "init_accuracy": 0.0,
      "n": 2.0,
      "op_accuracy": 1.0,
      "program_exact": 0.0,
      "repair_avg_candidates": 8.0,
      "repair_avg_verified": 0.5,
      "repair_changed_fraction": 0.0,
      "repair_executor_accuracy": 0.5,
      "repair_found_fraction": 0.5,
      "repair_program_exact": 0.0,
      "repair_state_accuracy": 0.5,
      "repair_state_prefix_fraction": 0.5,
      "state_accuracy": 0.5,
      "state_prefix_fraction": 0.5
    },
    "domain_boolean": {
      "arg_accuracy": 0.6666666666666666,
      "direct_accuracy": 0.5,
      "domain_boolean_direct_accuracy": 0.5,
      "domain_boolean_executor_accuracy": 1.0,
      "domain_boolean_n": 2.0,
      "domain_boolean_repair_executor_accuracy": 1.0,
      "executor_accuracy": 1.0,
      "init_accuracy": 0.5,
      "n": 2.0,
      "op_accuracy": 1.0,
      "program_exact": 0.0,
      "repair_avg_candidates": 8.0,
      "repair_avg_verified": 1.0,
      "repair_changed_fraction": 0.0,
      "repair_executor_accuracy": 1.0,
      "repair_found_fraction": 1.0,
      "repair_program_exact": 0.0,
      "repair_state_accuracy": 0.3333333333333333,
      "repair_state_prefix_fraction": 0.0,
      "state_accuracy": 0.3333333333333333,
      "state_prefix_fraction": 0.0
    },
    "domain_calendar": {
      "arg_accuracy": 0.16666666666666666,
      "direct_accuracy": 0.0,
      "domain_calendar_direct_accuracy": 0.0,
      "domain_calendar_executor_accuracy": 0.0,
      "domain_calendar_n": 2.0,
      "domain_calendar_repair_executor_accuracy": 1.0,
      "executor_accuracy": 0.0,
      "init_accuracy": 0.0,
      "n": 2.0,
      "op_accuracy": 1.0,
      "program_exact": 0.0,
      "repair_avg_candidates": 8.0,
      "repair_avg_verified": 1.0,
      "repair_changed_fraction": 1.0,
      "repair_executor_accuracy": 1.0,
      "repair_found_fraction": 1.0,
      "repair_program_exact": 0.0,
      "repair_state_accuracy": 0.5,
      "repair_state_prefix_fraction": 0.0,
      "state_accuracy": 0.0,
      "state_prefix_fraction": 0.0
    },
    "domain_list": {
      "arg_accuracy": 0.5,
      "direct_accuracy": 0.0,
      "domain_list_direct_accuracy": 0.0,
      "domain_list_executor_accuracy": 0.5,
      "domain_list_n": 2.0,
      "domain_list_repair_executor_accuracy": 1.0,
      "executor_accuracy": 0.5,
      "init_accuracy": 0.0,
      "n": 2.0,
      "op_accuracy": 1.0,
      "program_exact": 0.0,
      "repair_avg_candidates": 8.0,
      "repair_avg_verified": 1.0,
      "repair_changed_fraction": 0.5,
      "repair_executor_accuracy": 1.0,
      "repair_found_fraction": 1.0,
      "repair_program_exact": 0.0,
      "repair_state_accuracy": 1.0,
      "repair_state_prefix_fraction": 1.0,
      "state_accuracy": 0.6666666666666666,
      "state_prefix_fraction": 0.6666666666666666
    },
    "domain_lookup": {
      "arg_accuracy": 0.16666666666666666,
      "direct_accuracy": 0.0,
      "domain_lookup_direct_accuracy": 0.0,
      "domain_lookup_executor_accuracy": 0.0,
      "domain_lookup_n": 2.0,
      "domain_lookup_repair_executor_accuracy": 0.0,
      "executor_accuracy": 0.0,
      "init_accuracy": 0.0,
      "n": 2.0,
      "op_accuracy": 0.3333333333333333,
      "program_exact": 0.0,
      "repair_avg_candidates": 8.0,
      "repair_avg_verified": 0.0,
      "repair_changed_fraction": 0.0,
      "repair_executor_accuracy": 0.0,
      "repair_found_fraction": 0.0,
      "repair_program_exact": 0.0,
      "repair_state_accuracy": 0.16666666666666666,
      "repair_state_prefix_fraction": 0.16666666666666666,
      "state_accuracy": 0.16666666666666666,
      "state_prefix_fraction": 0.16666666666666666
    },
    "domain_unit": {
      "arg_accuracy": 0.5,
      "direct_accuracy": 0.0,
      "domain_unit_direct_accuracy": 0.0,
      "domain_unit_executor_accuracy": 0.0,
      "domain_unit_n": 2.0,
      "domain_unit_repair_executor_accuracy": 0.5,
      "executor_accuracy": 0.0,
      "init_accuracy": 0.0,
      "n": 2.0,
      "op_accuracy": 0.8333333333333334,
      "program_exact": 0.0,
      "repair_avg_candidates": 8.0,
      "repair_avg_verified": 0.5,
      "repair_changed_fraction": 0.5,
      "repair_executor_accuracy": 0.5,
      "repair_found_fraction": 0.5,
      "repair_program_exact": 0.0,
      "repair_state_accuracy": 0.5,
      "repair_state_prefix_fraction": 0.5,
      "state_accuracy": 0.16666666666666666,
      "state_prefix_fraction": 0.16666666666666666
    },
    "fresh_paired_mixed": {
      "arg_accuracy": 0.375,
      "compiler_pair_program_consistency": 0.25,
      "compiler_pair_state_consistency": 0.25,
      "direct_accuracy": 0.125,
      "domain_arithmetic_direct_accuracy": 0.0,
      "domain_arithmetic_executor_accuracy": 0.0,
      "domain_arithmetic_n": 2.0,
      "domain_arithmetic_repair_executor_accuracy": 0.0,
      "domain_calendar_direct_accuracy": 0.5,
      "domain_calendar_executor_accuracy": 0.0,
      "domain_calendar_n": 2.0,
      "domain_calendar_repair_executor_accuracy": 1.0,
      "domain_list_direct_accuracy": 0.0,
      "domain_list_executor_accuracy": 0.5,
      "domain_list_n": 2.0,
      "domain_list_repair_executor_accuracy": 1.0,
      "domain_unit_direct_accuracy": 0.0,
      "domain_unit_executor_accuracy": 0.5,
      "domain_unit_n": 2.0,
      "domain_unit_repair_executor_accuracy": 0.5,
      "executor_accuracy": 0.25,
      "executor_pair_answer_consistency": 0.25,
      "executor_pair_both_correct": 0.0,
      "init_accuracy": 0.25,
      "n": 8.0,
      "op_accuracy": 0.875,
      "program_exact": 0.0,
      "repair_avg_candidates": 8.0,
      "repair_avg_verified": 0.625,
      "repair_changed_fraction": 0.375,
      "repair_executor_accuracy": 0.625,
      "repair_found_fraction": 0.625,
      "repair_pair_answer_consistency": 0.5,
      "repair_pair_both_correct": 0.5,
      "repair_pair_state_consistency": 0.5,
      "repair_program_exact": 0.0,
      "repair_state_accuracy": 0.5,
      "repair_state_prefix_fraction": 0.4166666666666667,
      "state_accuracy": 0.2916666666666667,
      "state_prefix_fraction": 0.2916666666666667
    },
    "fresh_paraphrase_mixed": {
      "arg_accuracy": 0.6111111111111112,
      "direct_accuracy": 0.0,
      "domain_arithmetic_direct_accuracy": 0.0,
      "domain_arithmetic_executor_accuracy": 0.0,
      "domain_arithmetic_n": 1.0,
      "domain_arithmetic_repair_executor_accuracy": 1.0,
      "domain_boolean_direct_accuracy": 0.0,
      "domain_boolean_executor_accuracy": 1.0,
      "domain_boolean_n": 1.0,
      "domain_boolean_repair_executor_accuracy": 1.0,
      "domain_calendar_direct_accuracy": 0.0,
      "domain_calendar_executor_accuracy": 0.0,
      "domain_calendar_n": 1.0,
      "domain_calendar_repair_executor_accuracy": 1.0,
      "domain_list_direct_accuracy": 0.0,
      "domain_list_executor_accuracy": 1.0,
      "domain_list_n": 1.0,
      "domain_list_repair_executor_accuracy": 1.0,
      "domain_lookup_direct_accuracy": 0.0,
      "domain_lookup_executor_accuracy": 0.0,
      "domain_lookup_n": 1.0,
      "domain_lookup_repair_executor_accuracy": 0.0,
      "domain_unit_direct_accuracy": 0.0,
      "domain_unit_executor_accuracy": 0.0,
      "domain_unit_n": 1.0,
      "domain_unit_repair_executor_accuracy": 0.0,
      "executor_accuracy": 0.3333333333333333,
      "init_accuracy": 0.3333333333333333,
      "n": 6.0,
      "op_accuracy": 0.8333333333333334,
      "program_exact": 0.0,
      "repair_avg_candidates": 8.0,
      "repair_avg_verified": 0.6666666666666666,
      "repair_changed_fraction": 0.3333333333333333,
      "repair_executor_accuracy": 0.6666666666666666,
      "repair_found_fraction": 0.6666666666666666,
      "repair_program_exact": 0.0,
      "repair_state_accuracy": 0.5,
      "repair_state_prefix_fraction": 0.3888888888888889,
      "state_accuracy": 0.2777777777777778,
      "state_prefix_fraction": 0.2222222222222222
    },
    "fresh_standard_mixed": {
      "arg_accuracy": 0.4444444444444444,
      "direct_accuracy": 0.16666666666666666,
      "domain_arithmetic_direct_accuracy": 0.0,
      "domain_arithmetic_executor_accuracy": 0.0,
      "domain_arithmetic_n": 1.0,
      "domain_arithmetic_repair_executor_accuracy": 0.0,
      "domain_boolean_direct_accuracy": 1.0,
      "domain_boolean_executor_accuracy": 1.0,
      "domain_boolean_n": 1.0,
      "domain_boolean_repair_executor_accuracy": 1.0,
      "domain_calendar_direct_accuracy": 0.0,
      "domain_calendar_executor_accuracy": 0.0,
      "domain_calendar_n": 1.0,
      "domain_calendar_repair_executor_accuracy": 1.0,
      "domain_list_direct_accuracy": 0.0,
      "domain_list_executor_accuracy": 1.0,
      "domain_list_n": 1.0,
      "domain_list_repair_executor_accuracy": 1.0,
      "domain_lookup_direct_accuracy": 0.0,
      "domain_lookup_executor_accuracy": 0.0,
      "domain_lookup_n": 1.0,
      "domain_lookup_repair_executor_accuracy": 0.0,
      "domain_unit_direct_accuracy": 0.0,
      "domain_unit_executor_accuracy": 0.0,
      "domain_unit_n": 1.0,
      "domain_unit_repair_executor_accuracy": 0.0,
      "executor_accuracy": 0.3333333333333333,
      "init_accuracy": 0.6666666666666666,
      "n": 6.0,
      "op_accuracy": 0.7777777777777778,
      "program_exact": 0.0,
      "repair_avg_candidates": 8.0,
      "repair_avg_verified": 0.5,
      "repair_changed_fraction": 0.16666666666666666,
      "repair_executor_accuracy": 0.5,
      "repair_found_fraction": 0.5,
      "repair_program_exact": 0.0,
      "repair_state_accuracy": 0.2777777777777778,
      "repair_state_prefix_fraction": 0.16666666666666666,
      "state_accuracy": 0.2222222222222222,
      "state_prefix_fraction": 0.16666666666666666
    },
    "hard_paraphrase_mixed": {
      "arg_accuracy": 0.5833333333333334,
      "direct_accuracy": 0.3333333333333333,
      "domain_arithmetic_direct_accuracy": 0.0,
      "domain_arithmetic_executor_accuracy": 0.0,
      "domain_arithmetic_n": 1.0,
      "domain_arithmetic_repair_executor_accuracy": 0.0,
      "domain_boolean_direct_accuracy": 1.0,
      "domain_boolean_executor_accuracy": 0.0,
      "domain_boolean_n": 1.0,
      "domain_boolean_repair_executor_accuracy": 1.0,
      "domain_calendar_direct_accuracy": 1.0,
      "domain_calendar_executor_accuracy": 0.0,
      "domain_calendar_n": 1.0,
      "domain_calendar_repair_executor_accuracy": 1.0,
      "domain_list_direct_accuracy": 0.0,
      "domain_list_executor_accuracy": 0.0,
      "domain_list_n": 1.0,
      "domain_list_repair_executor_accuracy": 0.0,
      "domain_lookup_direct_accuracy": 0.0,
      "domain_lookup_executor_accuracy": 0.0,
      "domain_lookup_n": 1.0,
      "domain_lookup_repair_executor_accuracy": 0.0,
      "domain_unit_direct_accuracy": 0.0,
      "domain_unit_executor_accuracy": 0.0,
      "domain_unit_n": 1.0,
      "domain_unit_repair_executor_accuracy": 1.0,
      "executor_accuracy": 0.0,
      "init_accuracy": 0.0,
      "n": 6.0,
      "op_accuracy": 0.75,
      "program_exact": 0.0,
      "repair_avg_candidates": 10.0,
      "repair_avg_verified": 0.5,
      "repair_changed_fraction": 0.5,
      "repair_executor_accuracy": 0.5,
      "repair_found_fraction": 0.5,
      "repair_program_exact": 0.0,
      "repair_state_accuracy": 0.2916666666666667,
      "repair_state_prefix_fraction": 0.16666666666666666,
      "state_accuracy": 0.08333333333333333,
      "state_prefix_fraction": 0.041666666666666664
    },
    "hard_standard_mixed": {
      "arg_accuracy": 0.3333333333333333,
      "direct_accuracy": 0.0,
      "domain_arithmetic_direct_accuracy": 0.0,
      "domain_arithmetic_executor_accuracy": 0.0,
      "domain_arithmetic_n": 1.0,
      "domain_arithmetic_repair_executor_accuracy": 0.0,
      "domain_boolean_direct_accuracy": 0.0,
      "domain_boolean_executor_accuracy": 1.0,
      "domain_boolean_n": 1.0,
      "domain_boolean_repair_executor_accuracy": 1.0,
      "domain_calendar_direct_accuracy": 0.0,
      "domain_calendar_executor_accuracy": 0.0,
      "domain_calendar_n": 1.0,
      "domain_calendar_repair_executor_accuracy": 1.0,
      "domain_list_direct_accuracy": 0.0,
      "domain_list_executor_accuracy": 1.0,
      "domain_list_n": 1.0,
      "domain_list_repair_executor_accuracy": 1.0,
      "domain_lookup_direct_accuracy": 0.0,
      "domain_lookup_executor_accuracy": 0.0,
      "domain_lookup_n": 1.0,
      "domain_lookup_repair_executor_accuracy": 1.0,
      "domain_unit_direct_accuracy": 0.0,
      "domain_unit_executor_accuracy": 0.0,
      "domain_unit_n": 1.0,
      "domain_unit_repair_executor_accuracy": 0.0,
      "executor_accuracy": 0.3333333333333333,
      "init_accuracy": 0.3333333333333333,
      "n": 6.0,
      "op_accuracy": 0.7083333333333334,
      "program_exact": 0.0,
      "repair_avg_candidates": 10.0,
      "repair_avg_verified": 0.6666666666666666,
      "repair_changed_fraction": 0.3333333333333333,
      "repair_executor_accuracy": 0.6666666666666666,
      "repair_found_fraction": 0.6666666666666666,
      "repair_program_exact": 0.0,
      "repair_state_accuracy": 0.2916666666666667,
      "repair_state_prefix_fraction": 0.041666666666666664,
      "state_accuracy": 0.25,
      "state_prefix_fraction": 0.041666666666666664
    },
    "val_mixed": {
      "arg_accuracy": 0.3888888888888889,
      "direct_accuracy": 0.0,
      "domain_arithmetic_direct_accuracy": 0.0,
      "domain_arithmetic_executor_accuracy": 0.0,
      "domain_arithmetic_n": 1.0,
      "domain_arithmetic_repair_executor_accuracy": 0.0,
      "domain_boolean_direct_accuracy": 0.0,
      "domain_boolean_executor_accuracy": 1.0,
      "domain_boolean_n": 1.0,
      "domain_boolean_repair_executor_accuracy": 1.0,
      "domain_calendar_direct_accuracy": 0.0,
      "domain_calendar_executor_accuracy": 1.0,
      "domain_calendar_n": 1.0,
      "domain_calendar_repair_executor_accuracy": 1.0,
      "domain_list_direct_accuracy": 0.0,
      "domain_list_executor_accuracy": 0.0,
      "domain_list_n": 1.0,
      "domain_list_repair_executor_accuracy": 1.0,
      "domain_lookup_direct_accuracy": 0.0,
      "domain_lookup_executor_accuracy": 0.0,
      "domain_lookup_n": 1.0,
      "domain_lookup_repair_executor_accuracy": 0.0,
      "domain_unit_direct_accuracy": 0.0,
      "domain_unit_executor_accuracy": 0.0,
      "domain_unit_n": 1.0,
      "domain_unit_repair_executor_accuracy": 0.0,
      "executor_accuracy": 0.3333333333333333,
      "init_accuracy": 0.3333333333333333,
      "n": 6.0,
      "op_accuracy": 0.8333333333333334,
      "program_exact": 0.0,
      "repair_avg_candidates": 8.0,
      "repair_avg_verified": 0.5,
      "repair_changed_fraction": 0.16666666666666666,
      "repair_executor_accuracy": 0.5,
      "repair_found_fraction": 0.5,
      "repair_program_exact": 0.0,
      "repair_state_accuracy": 0.2777777777777778,
      "repair_state_prefix_fraction": 0.16666666666666666,
      "state_accuracy": 0.1111111111111111,
      "state_prefix_fraction": 0.0
    }
  },
  "train_log": [
    {
      "domain_arithmetic_arg_accuracy": 0.16666666666666666,
      "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": 2.0,
      "domain_arithmetic_domain_arithmetic_repair_executor_accuracy": 0.5,
      "domain_arithmetic_executor_accuracy": 0.0,
      "domain_arithmetic_init_accuracy": 0.0,
      "domain_arithmetic_op_accuracy": 1.0,
      "domain_arithmetic_program_exact": 0.0,
      "domain_arithmetic_repair_avg_candidates": 8.0,
      "domain_arithmetic_repair_avg_verified": 0.5,
      "domain_arithmetic_repair_changed_fraction": 0.5,
      "domain_arithmetic_repair_executor_accuracy": 0.5,
      "domain_arithmetic_repair_found_fraction": 0.5,
      "domain_arithmetic_repair_program_exact": 0.0,
      "domain_arithmetic_repair_state_accuracy": 0.5,
      "domain_arithmetic_repair_state_prefix_fraction": 0.5,
      "domain_arithmetic_state_accuracy": 0.16666666666666666,
      "domain_arithmetic_state_prefix_fraction": 0.16666666666666666,
      "domain_boolean_arg_accuracy": 0.5,
      "domain_boolean_direct_accuracy": 0.5,
      "domain_boolean_domain_boolean_direct_accuracy": 0.5,
      "domain_boolean_domain_boolean_executor_accuracy": 0.5,
      "domain_boolean_domain_boolean_n": 2.0,
      "domain_boolean_domain_boolean_repair_executor_accuracy": 1.0,
      "domain_boolean_executor_accuracy": 0.5,
      "domain_boolean_init_accuracy": 0.0,
      "domain_boolean_op_accuracy": 0.6666666666666666,
      "domain_boolean_program_exact": 0.0,
      "domain_boolean_repair_avg_candidates": 8.0,
      "domain_boolean_repair_avg_verified": 1.0,
      "domain_boolean_repair_changed_fraction": 0.5,
      "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": 0.3333333333333333,
      "domain_boolean_repair_state_prefix_fraction": 0.0,
      "domain_boolean_state_accuracy": 0.16666666666666666,
      "domain_boolean_state_prefix_fraction": 0.0,
      "domain_calendar_arg_accuracy": 0.16666666666666666,
      "domain_calendar_direct_accuracy": 0.0,
      "domain_calendar_domain_calendar_direct_accuracy": 0.0,
      "domain_calendar_domain_calendar_executor_accuracy": 0.5,
      "domain_calendar_domain_calendar_n": 2.0,
      "domain_calendar_domain_calendar_repair_executor_accuracy": 1.0,
      "domain_calendar_executor_accuracy": 0.5,
      "domain_calendar_init_accuracy": 0.0,
      "domain_calendar_op_accuracy": 1.0,
      "domain_calendar_program_exact": 0.0,
      "domain_calendar_repair_avg_candidates": 8.0,
      "domain_calendar_repair_avg_verified": 1.0,
      "domain_calendar_repair_changed_fraction": 0.5,
      "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": 0.5,
      "domain_calendar_repair_state_prefix_fraction": 0.0,
      "domain_calendar_state_accuracy": 0.3333333333333333,
      "domain_calendar_state_prefix_fraction": 0.0,
      "domain_list_arg_accuracy": 0.16666666666666666,
      "domain_list_direct_accuracy": 0.0,
      "domain_list_domain_list_direct_accuracy": 0.0,
      "domain_list_domain_list_executor_accuracy": 1.0,
      "domain_list_domain_list_n": 2.0,
      "domain_list_domain_list_repair_executor_accuracy": 1.0,
      "domain_list_executor_accuracy": 1.0,
      "domain_list_init_accuracy": 0.0,
      "domain_list_op_accuracy": 0.8333333333333334,
      "domain_list_program_exact": 0.0,
      "domain_list_repair_avg_candidates": 8.0,
      "domain_list_repair_avg_verified": 1.0,
      "domain_list_repair_changed_fraction": 0.0,
      "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": 0.8333333333333334,
      "domain_list_repair_state_prefix_fraction": 0.6666666666666666,
      "domain_list_state_accuracy": 0.8333333333333334,
      "domain_list_state_prefix_fraction": 0.6666666666666666,
      "domain_lookup_arg_accuracy": 0.16666666666666666,
      "domain_lookup_direct_accuracy": 0.0,
      "domain_lookup_domain_lookup_direct_accuracy": 0.0,
      "domain_lookup_domain_lookup_executor_accuracy": 0.0,
      "domain_lookup_domain_lookup_n": 2.0,
      "domain_lookup_domain_lookup_repair_executor_accuracy": 0.0,
      "domain_lookup_executor_accuracy": 0.0,
      "domain_lookup_init_accuracy": 0.0,
      "domain_lookup_op_accuracy": 0.3333333333333333,
      "domain_lookup_program_exact": 0.0,
      "domain_lookup_repair_avg_candidates": 8.0,
      "domain_lookup_repair_avg_verified": 0.0,
      "domain_lookup_repair_changed_fraction": 0.0,
      "domain_lookup_repair_executor_accuracy": 0.0,
      "domain_lookup_repair_found_fraction": 0.0,
      "domain_lookup_repair_program_exact": 0.0,
      "domain_lookup_repair_state_accuracy": 0.16666666666666666,
      "domain_lookup_repair_state_prefix_fraction": 0.16666666666666666,
      "domain_lookup_state_accuracy": 0.16666666666666666,
      "domain_lookup_state_prefix_fraction": 0.16666666666666666,
      "domain_unit_arg_accuracy": 0.3333333333333333,
      "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": 2.0,
      "domain_unit_domain_unit_repair_executor_accuracy": 0.0,
      "domain_unit_executor_accuracy": 0.0,
      "domain_unit_init_accuracy": 0.0,
      "domain_unit_op_accuracy": 0.8333333333333334,
      "domain_unit_program_exact": 0.0,
      "domain_unit_repair_avg_candidates": 8.0,
      "domain_unit_repair_avg_verified": 0.0,
      "domain_unit_repair_changed_fraction": 0.0,
      "domain_unit_repair_executor_accuracy": 0.0,
      "domain_unit_repair_found_fraction": 0.0,
      "domain_unit_repair_program_exact": 0.0,
      "domain_unit_repair_state_accuracy": 0.0,
      "domain_unit_repair_state_prefix_fraction": 0.0,
      "domain_unit_state_accuracy": 0.0,
      "domain_unit_state_prefix_fraction": 0.0,
      "fresh_paired_mixed_arg_accuracy": 0.125,
      "fresh_paired_mixed_compiler_pair_program_consistency": 0.0,
      "fresh_paired_mixed_compiler_pair_state_consistency": 0.25,
      "fresh_paired_mixed_direct_accuracy": 0.125,
      "fresh_paired_mixed_domain_arithmetic_direct_accuracy": 0.0,
      "fresh_paired_mixed_domain_arithmetic_executor_accuracy": 0.0,
      "fresh_paired_mixed_domain_arithmetic_n": 2.0,
      "fresh_paired_mixed_domain_arithmetic_repair_executor_accuracy": 0.0,
      "fresh_paired_mixed_domain_calendar_direct_accuracy": 0.5,
      "fresh_paired_mixed_domain_calendar_executor_accuracy": 0.0,
      "fresh_paired_mixed_domain_calendar_n": 2.0,
      "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": 1.0,
      "fresh_paired_mixed_domain_list_n": 2.0,
      "fresh_paired_mixed_domain_list_repair_executor_accuracy": 1.0,
      "fresh_paired_mixed_domain_unit_direct_accuracy": 0.0,
      "fresh_paired_mixed_domain_unit_executor_accuracy": 0.0,
      "fresh_paired_mixed_domain_unit_n": 2.0,
      "fresh_paired_mixed_domain_unit_repair_executor_accuracy": 0.0,
      "fresh_paired_mixed_executor_accuracy": 0.25,
      "fresh_paired_mixed_executor_pair_answer_consistency": 0.25,
      "fresh_paired_mixed_executor_pair_both_correct": 0.25,
      "fresh_paired_mixed_init_accuracy": 0.125,
      "fresh_paired_mixed_op_accuracy": 0.75,
      "fresh_paired_mixed_program_exact": 0.0,
      "fresh_paired_mixed_repair_avg_candidates": 8.0,
      "fresh_paired_mixed_repair_avg_verified": 0.5,
      "fresh_paired_mixed_repair_changed_fraction": 0.25,
      "fresh_paired_mixed_repair_executor_accuracy": 0.5,
      "fresh_paired_mixed_repair_found_fraction": 0.5,
      "fresh_paired_mixed_repair_pair_answer_consistency": 0.5,
      "fresh_paired_mixed_repair_pair_both_correct": 0.5,
      "fresh_paired_mixed_repair_pair_state_consistency": 0.25,
      "fresh_paired_mixed_repair_program_exact": 0.0,
      "fresh_paired_mixed_repair_state_accuracy": 0.5,
      "fresh_paired_mixed_repair_state_prefix_fraction": 0.4166666666666667,
      "fresh_paired_mixed_state_accuracy": 0.375,
      "fresh_paired_mixed_state_prefix_fraction": 0.375,
      "fresh_paraphrase_mixed_arg_accuracy": 0.4444444444444444,
      "fresh_paraphrase_mixed_direct_accuracy": 0.0,
      "fresh_paraphrase_mixed_domain_arithmetic_direct_accuracy": 0.0,
      "fresh_paraphrase_mixed_domain_arithmetic_executor_accuracy": 1.0,
      "fresh_paraphrase_mixed_domain_arithmetic_n": 1.0,
      "fresh_paraphrase_mixed_domain_arithmetic_repair_executor_accuracy": 1.0,
      "fresh_paraphrase_mixed_domain_boolean_direct_accuracy": 0.0,
      "fresh_paraphrase_mixed_domain_boolean_executor_accuracy": 1.0,
      "fresh_paraphrase_mixed_domain_boolean_n": 1.0,
      "fresh_paraphrase_mixed_domain_boolean_repair_executor_accuracy": 1.0,
      "fresh_paraphrase_mixed_domain_calendar_direct_accuracy": 0.0,
      "fresh_paraphrase_mixed_domain_calendar_executor_accuracy": 0.0,
      "fresh_paraphrase_mixed_domain_calendar_n": 1.0,
      "fresh_paraphrase_mixed_domain_calendar_repair_executor_accuracy": 0.0,
      "fresh_paraphrase_mixed_domain_list_direct_accuracy": 0.0,
      "fresh_paraphrase_mixed_domain_list_executor_accuracy": 1.0,
      "fresh_paraphrase_mixed_domain_list_n": 1.0,
      "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": 0.0,
      "fresh_paraphrase_mixed_domain_lookup_n": 1.0,
      "fresh_paraphrase_mixed_domain_lookup_repair_executor_accuracy": 0.0,
      "fresh_paraphrase_mixed_domain_unit_direct_accuracy": 0.0,
      "fresh_paraphrase_mixed_domain_unit_executor_accuracy": 0.0,
      "fresh_paraphrase_mixed_domain_unit_n": 1.0,
      "fresh_paraphrase_mixed_domain_unit_repair_executor_accuracy": 0.0,
      "fresh_paraphrase_mixed_executor_accuracy": 0.5,
      "fresh_paraphrase_mixed_init_accuracy": 0.16666666666666666,
      "fresh_paraphrase_mixed_op_accuracy": 0.7222222222222222,
      "fresh_paraphrase_mixed_program_exact": 0.0,
      "fresh_paraphrase_mixed_repair_avg_candidates": 8.0,
      "fresh_paraphrase_mixed_repair_avg_verified": 0.5,
      "fresh_paraphrase_mixed_repair_changed_fraction": 0.0,
      "fresh_paraphrase_mixed_repair_executor_accuracy": 0.5,
      "fresh_paraphrase_mixed_repair_found_fraction": 0.5,
      "fresh_paraphrase_mixed_repair_program_exact": 0.0,
      "fresh_paraphrase_mixed_repair_state_accuracy": 0.4444444444444444,
      "fresh_paraphrase_mixed_repair_state_prefix_fraction": 0.3888888888888889,
      "fresh_paraphrase_mixed_state_accuracy": 0.4444444444444444,
      "fresh_paraphrase_mixed_state_prefix_fraction": 0.3888888888888889,
      "fresh_standard_mixed_arg_accuracy": 0.2222222222222222,
      "fresh_standard_mixed_direct_accuracy": 0.16666666666666666,
      "fresh_standard_mixed_domain_arithmetic_direct_accuracy": 0.0,
      "fresh_standard_mixed_domain_arithmetic_executor_accuracy": 0.0,
      "fresh_standard_mixed_domain_arithmetic_n": 1.0,
      "fresh_standard_mixed_domain_arithmetic_repair_executor_accuracy": 0.0,
      "fresh_standard_mixed_domain_boolean_direct_accuracy": 1.0,
      "fresh_standard_mixed_domain_boolean_executor_accuracy": 1.0,
      "fresh_standard_mixed_domain_boolean_n": 1.0,
      "fresh_standard_mixed_domain_boolean_repair_executor_accuracy": 1.0,
      "fresh_standard_mixed_domain_calendar_direct_accuracy": 0.0,
      "fresh_standard_mixed_domain_calendar_executor_accuracy": 0.0,
      "fresh_standard_mixed_domain_calendar_n": 1.0,
      "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": 1.0,
      "fresh_standard_mixed_domain_list_n": 1.0,
      "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": 1.0,
      "fresh_standard_mixed_domain_lookup_repair_executor_accuracy": 0.0,
      "fresh_standard_mixed_domain_unit_direct_accuracy": 0.0,
      "fresh_standard_mixed_domain_unit_executor_accuracy": 0.0,
      "fresh_standard_mixed_domain_unit_n": 1.0,
      "fresh_standard_mixed_domain_unit_repair_executor_accuracy": 0.0,
      "fresh_standard_mixed_executor_accuracy": 0.3333333333333333,
      "fresh_standard_mixed_init_accuracy": 0.3333333333333333,
      "fresh_standard_mixed_op_accuracy": 0.6666666666666666,
      "fresh_standard_mixed_program_exact": 0.0,
      "fresh_standard_mixed_repair_avg_candidates": 8.0,
      "fresh_standard_mixed_repair_avg_verified": 0.5,
      "fresh_standard_mixed_repair_changed_fraction": 0.16666666666666666,
      "fresh_standard_mixed_repair_executor_accuracy": 0.5,
      "fresh_standard_mixed_repair_found_fraction": 0.5,
      "fresh_standard_mixed_repair_program_exact": 0.0,
      "fresh_standard_mixed_repair_state_accuracy": 0.2777777777777778,
      "fresh_standard_mixed_repair_state_prefix_fraction": 0.05555555555555555,
      "fresh_standard_mixed_state_accuracy": 0.2222222222222222,
      "fresh_standard_mixed_state_prefix_fraction": 0.05555555555555555,
      "hard_paraphrase_mixed_arg_accuracy": 0.4166666666666667,
      "hard_paraphrase_mixed_direct_accuracy": 0.3333333333333333,
      "hard_paraphrase_mixed_domain_arithmetic_direct_accuracy": 0.0,
      "hard_paraphrase_mixed_domain_arithmetic_executor_accuracy": 0.0,
      "hard_paraphrase_mixed_domain_arithmetic_n": 1.0,
      "hard_paraphrase_mixed_domain_arithmetic_repair_executor_accuracy": 0.0,
      "hard_paraphrase_mixed_domain_boolean_direct_accuracy": 1.0,
      "hard_paraphrase_mixed_domain_boolean_executor_accuracy": 0.0,
      "hard_paraphrase_mixed_domain_boolean_n": 1.0,
      "hard_paraphrase_mixed_domain_boolean_repair_executor_accuracy": 1.0,
      "hard_paraphrase_mixed_domain_calendar_direct_accuracy": 1.0,
      "hard_paraphrase_mixed_domain_calendar_executor_accuracy": 0.0,
      "hard_paraphrase_mixed_domain_calendar_n": 1.0,
      "hard_paraphrase_mixed_domain_calendar_repair_executor_accuracy": 0.0,
      "hard_paraphrase_mixed_domain_list_direct_accuracy": 0.0,
      "hard_paraphrase_mixed_domain_list_executor_accuracy": 0.0,
      "hard_paraphrase_mixed_domain_list_n": 1.0,
      "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": 1.0,
      "hard_paraphrase_mixed_domain_lookup_repair_executor_accuracy": 0.0,
      "hard_paraphrase_mixed_domain_unit_direct_accuracy": 0.0,
      "hard_paraphrase_mixed_domain_unit_executor_accuracy": 0.0,
      "hard_paraphrase_mixed_domain_unit_n": 1.0,
      "hard_paraphrase_mixed_domain_unit_repair_executor_accuracy": 0.0,
      "hard_paraphrase_mixed_executor_accuracy": 0.0,
      "hard_paraphrase_mixed_init_accuracy": 0.0,
      "hard_paraphrase_mixed_op_accuracy": 0.75,
      "hard_paraphrase_mixed_program_exact": 0.0,
      "hard_paraphrase_mixed_repair_avg_candidates": 10.0,
      "hard_paraphrase_mixed_repair_avg_verified": 0.3333333333333333,
      "hard_paraphrase_mixed_repair_changed_fraction": 0.3333333333333333,
      "hard_paraphrase_mixed_repair_executor_accuracy": 0.3333333333333333,
      "hard_paraphrase_mixed_repair_found_fraction": 0.3333333333333333,
      "hard_paraphrase_mixed_repair_program_exact": 0.0,
      "hard_paraphrase_mixed_repair_state_accuracy": 0.25,
      "hard_paraphrase_mixed_repair_state_prefix_fraction": 0.041666666666666664,
      "hard_paraphrase_mixed_state_accuracy": 0.08333333333333333,
      "hard_paraphrase_mixed_state_prefix_fraction": 0.041666666666666664,
      "hard_standard_mixed_arg_accuracy": 0.2916666666666667,
      "hard_standard_mixed_direct_accuracy": 0.0,
      "hard_standard_mixed_domain_arithmetic_direct_accuracy": 0.0,
      "hard_standard_mixed_domain_arithmetic_executor_accuracy": 0.0,
      "hard_standard_mixed_domain_arithmetic_n": 1.0,
      "hard_standard_mixed_domain_arithmetic_repair_executor_accuracy": 0.0,
      "hard_standard_mixed_domain_boolean_direct_accuracy": 0.0,
      "hard_standard_mixed_domain_boolean_executor_accuracy": 0.0,
      "hard_standard_mixed_domain_boolean_n": 1.0,
      "hard_standard_mixed_domain_boolean_repair_executor_accuracy": 1.0,
      "hard_standard_mixed_domain_calendar_direct_accuracy": 0.0,
      "hard_standard_mixed_domain_calendar_executor_accuracy": 0.0,
      "hard_standard_mixed_domain_calendar_n": 1.0,
      "hard_standard_mixed_domain_calendar_repair_executor_accuracy": 1.0,
      "hard_standard_mixed_domain_list_direct_accuracy": 0.0,
      "hard_standard_mixed_domain_list_executor_accuracy": 1.0,
      "hard_standard_mixed_domain_list_n": 1.0,
      "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": 0.0,
      "hard_standard_mixed_domain_lookup_n": 1.0,
      "hard_standard_mixed_domain_lookup_repair_executor_accuracy": 0.0,
      "hard_standard_mixed_domain_unit_direct_accuracy": 0.0,
      "hard_standard_mixed_domain_unit_executor_accuracy": 0.0,
      "hard_standard_mixed_domain_unit_n": 1.0,
      "hard_standard_mixed_domain_unit_repair_executor_accuracy": 0.0,
      "hard_standard_mixed_executor_accuracy": 0.16666666666666666,
      "hard_standard_mixed_init_accuracy": 0.0,
      "hard_standard_mixed_op_accuracy": 0.7083333333333334,
      "hard_standard_mixed_program_exact": 0.0,
      "hard_standard_mixed_repair_avg_candidates": 10.0,
      "hard_standard_mixed_repair_avg_verified": 0.5,
      "hard_standard_mixed_repair_changed_fraction": 0.3333333333333333,
      "hard_standard_mixed_repair_executor_accuracy": 0.5,
      "hard_standard_mixed_repair_found_fraction": 0.5,
      "hard_standard_mixed_repair_program_exact": 0.0,
      "hard_standard_mixed_repair_state_accuracy": 0.2916666666666667,
      "hard_standard_mixed_repair_state_prefix_fraction": 0.20833333333333334,
      "hard_standard_mixed_state_accuracy": 0.16666666666666666,
      "hard_standard_mixed_state_prefix_fraction": 0.16666666666666666,
      "phase": "baseline",
      "step": 0,
      "val_mixed_arg_accuracy": 0.2222222222222222,
      "val_mixed_direct_accuracy": 0.0,
      "val_mixed_domain_arithmetic_direct_accuracy": 0.0,
      "val_mixed_domain_arithmetic_executor_accuracy": 0.0,
      "val_mixed_domain_arithmetic_n": 1.0,
      "val_mixed_domain_arithmetic_repair_executor_accuracy": 0.0,
      "val_mixed_domain_boolean_direct_accuracy": 0.0,
      "val_mixed_domain_boolean_executor_accuracy": 1.0,
      "val_mixed_domain_boolean_n": 1.0,
      "val_mixed_domain_boolean_repair_executor_accuracy": 1.0,
      "val_mixed_domain_calendar_direct_accuracy": 0.0,
      "val_mixed_domain_calendar_executor_accuracy": 0.0,
      "val_mixed_domain_calendar_n": 1.0,
      "val_mixed_domain_calendar_repair_executor_accuracy": 0.0,
      "val_mixed_domain_list_direct_accuracy": 0.0,
      "val_mixed_domain_list_executor_accuracy": 0.0,
      "val_mixed_domain_list_n": 1.0,
      "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": 1.0,
      "val_mixed_domain_lookup_repair_executor_accuracy": 0.0,
      "val_mixed_domain_unit_direct_accuracy": 0.0,
      "val_mixed_domain_unit_executor_accuracy": 0.0,
      "val_mixed_domain_unit_n": 1.0,
      "val_mixed_domain_unit_repair_executor_accuracy": 1.0,
      "val_mixed_executor_accuracy": 0.16666666666666666,
      "val_mixed_init_accuracy": 0.16666666666666666,
      "val_mixed_op_accuracy": 0.6666666666666666,
      "val_mixed_program_exact": 0.0,
      "val_mixed_repair_avg_candidates": 8.0,
      "val_mixed_repair_avg_verified": 0.5,
      "val_mixed_repair_changed_fraction": 0.3333333333333333,
      "val_mixed_repair_executor_accuracy": 0.5,
      "val_mixed_repair_found_fraction": 0.5,
      "val_mixed_repair_program_exact": 0.0,
      "val_mixed_repair_state_accuracy": 0.2777777777777778,
      "val_mixed_repair_state_prefix_fraction": 0.16666666666666666,
      "val_mixed_state_accuracy": 0.16666666666666666,
      "val_mixed_state_prefix_fraction": 0.1111111111111111
    },
    {
      "arg_attn_loss": 3.8433802127838135,
      "arg_loss": 1.6396098136901855,
      "executor_loss": 0.9130217432975769,
      "init_attn_loss": 3.7053956985473633,
      "init_loss": 0.41356584429740906,
      "loss": 32.71820068359375,
      "op_attn_loss": 3.7622101306915283,
      "op_loss": 0.0,
      "phase": "curriculum_train",
      "role_loss": 1.4361145496368408,
      "stage_max_len": 2,
      "state_loss": 0.9130217432975769,
      "state_train_accuracy": 0.0,
      "step": 1
    },
    {
      "arg_attn_loss": 3.950740337371826,
      "arg_loss": 1.1505320072174072,
      "executor_loss": 1.869184970855713,
      "init_attn_loss": 3.711458921432495,
      "init_loss": 0.7186530828475952,
      "loss": 32.703025817871094,
      "op_attn_loss": 3.8435356616973877,
      "op_loss": 0.0,
      "phase": "curriculum_train",
      "role_loss": 1.4010316133499146,
      "stage_max_len": 3,
      "state_loss": 1.869184970855713,
      "state_train_accuracy": 0.0,
      "step": 2
    },
    {
      "phase": "repair_targets",
      "repair_avg_candidates": 5.0,
      "repair_avg_verified": 0.5,
      "repair_changed_fraction": 0.3333333333333333,
      "repair_fallback_gold_fraction": 0.0,
      "repair_found_fraction": 0.5,
      "repair_source_n": 12.0,
      "repair_train_n": 4.0,
      "step": 2
    },
    {
      "arg_loss": 0.42146262526512146,
      "executor_loss": 0.40498480200767517,
      "init_loss": 0.7084128260612488,
      "loss": 4.620748043060303,
      "op_loss": 0.0,
      "phase": "repair_distill",
      "repair_step": 1,
      "state_loss": 0.40498480200767517,
      "state_train_accuracy": 1.0,
      "step": 3
    }
  ]
}
