{
  "args": {
    "model_id": "Qwen/Qwen3-4B",
    "load_in_4bit": true,
    "torch_dtype": "bf16",
    "device_map": "auto",
    "use_lora": true,
    "lora_r": 8,
    "lora_alpha": 16,
    "lora_dropout": 0.05,
    "lora_target_modules": "all-linear",
    "gradient_checkpointing": true,
    "optimizer": "paged_adamw_8bit",
    "modulus": 97,
    "max_steps": 24,
    "train_min_len": 1,
    "train_max_len": 12,
    "curriculum_stages": "short:1:4:80,medium:1:8:80,train:1:12:80",
    "answer_train_min_len": 1,
    "answer_train_max_len": 24,
    "eval_lengths": "4,8,12,24",
    "train_template_mode": "mixed",
    "paired_train": true,
    "paired_batches": true,
    "paired_template_modes": "standard,paraphrase",
    "pair_size": 2,
    "answer_train_template_mode": "mixed",
    "eval_template_modes": "standard,paraphrase",
    "paired_eval": true,
    "paired_eval_template_modes": "standard,paraphrase",
    "train_size": 256,
    "answer_train_size": 256,
    "eval_size": 64,
    "train_batch_size": 4,
    "eval_batch_size": 8,
    "bootstrap_steps": 0,
    "answer_steps": 0,
    "eval_every": 80,
    "stage_eval_every": 80,
    "lr": 0.0002,
    "answer_lr_scale": 0.1,
    "weight_decay": 0.0,
    "grad_clip": 1.0,
    "head_width": 256,
    "register_width": 256,
    "register_layers": 1,
    "register_heads": 4,
    "register_dropout": 0.05,
    "register_style": "bare",
    "trace_loss_weight": 1.0,
    "init_trace_loss_weight": 4.0,
    "op_trace_loss_weight": 1.0,
    "arg_trace_loss_weight": 4.0,
    "executor_loss_weight": 1.0,
    "state_loss_weight": 1.0,
    "pair_consistency_loss_weight": 0.1,
    "pair_state_consistency_loss_weight": 0.5,
    "variants": "structured_trace_state_consistency",
    "max_length": 768,
    "seed": 7,
    "train_seed": -1,
    "output_dir": "experiments/qwen_register_structured_runtime/runs/pilot_structured_state_consistency_s240",
    "checkpoint_dir": "large_artifacts/qwen_register_structured_runtime/checkpoints/pilot_structured_state_consistency_s240"
  },
  "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,
    "transformers_loader": "AutoModelForCausalLM",
    "peft_installed": true,
    "use_lora": true,
    "lora_r": 8,
    "gpu_name": "NVIDIA RTX 6000 Ada Generation",
    "gpu_vram_gb": 47.383
  },
  "hidden_dim": 2560,
  "register_count": 49,
  "dataset": {
    "train_size": 256,
    "train_examples": 1536,
    "curriculum_stages": [
      {
        "name": "short",
        "min_len": 1,
        "max_len": 4,
        "steps": 80
      },
      {
        "name": "medium",
        "min_len": 1,
        "max_len": 8,
        "steps": 80
      },
      {
        "name": "train",
        "min_len": 1,
        "max_len": 12,
        "steps": 80
      }
    ],
    "answer_train_size": 256,
    "eval_size": 64,
    "eval_lengths": [
      4,
      8,
      12,
      24
    ],
    "train_template_mode": "mixed",
    "paired_train": true,
    "paired_template_modes": [
      "standard",
      "paraphrase"
    ],
    "answer_train_template_mode": "mixed",
    "eval_template_modes": [
      "standard",
      "paraphrase"
    ],
    "paired_eval": true,
    "paired_eval_template_modes": [
      "standard",
      "paraphrase"
    ],
    "register_style": "bare"
  },
  "variants": {
    "structured_trace_state_consistency": {
      "variant": "structured_trace_state_consistency",
      "train_seconds": 921.9288547039032,
      "checkpoints": [
        "large_artifacts/qwen_register_structured_runtime/checkpoints/pilot_structured_state_consistency_s240/structured_trace_state_consistency/adapter",
        "large_artifacts/qwen_register_structured_runtime/checkpoints/pilot_structured_state_consistency_s240/structured_trace_state_consistency/heads.pt"
      ],
      "train_log": [
        {
          "variant": "structured_trace_state_consistency",
          "stage": "structured_trace_state_consistency_short",
          "local_step": 1,
          "step": 1,
          "loss": 47.07431411743164,
          "trace_loss_active": true,
          "executor_loss_active": true,
          "state_loss_active": true,
          "lr_scale": 1.0,
          "executor_loss": 4.582798957824707,
          "state_loss": 4.581491947174072,
          "state_train_accuracy": 0.0,
          "pair_program_loss": 0.0052149249240756035,
          "pair_state_loss": 1.6290605344693176e-05,
          "init_loss": 4.597006320953369,
          "op_loss": 1.1381311416625977,
          "arg_loss": 4.595834255218506,
          "standard_len4_direct_accuracy": NaN,
          "standard_len4_direct_target_mass": NaN,
          "standard_len4_executor_accuracy": 0.046875,
          "standard_len4_executor_target_mass": 0.010264429613016546,
          "standard_len4_init_accuracy": 0.015625,
          "standard_len4_op_accuracy": 0.33984375,
          "standard_len4_arg_accuracy": 0.04296875,
          "standard_len4_program_exact": 0.0,
          "standard_len4_state_accuracy": 0.01171875,
          "standard_len4_state_all_exact": 0.0,
          "standard_len4_state_prefix_fraction": 0.0,
          "standard_len8_direct_accuracy": NaN,
          "standard_len8_direct_target_mass": NaN,
          "standard_len8_executor_accuracy": 0.03125,
          "standard_len8_executor_target_mass": 0.010334888123907149,
          "standard_len8_init_accuracy": 0.015625,
          "standard_len8_op_accuracy": 0.322265625,
          "standard_len8_arg_accuracy": 0.029296875,
          "standard_len8_program_exact": 0.0,
          "standard_len8_state_accuracy": 0.013671875,
          "standard_len8_state_all_exact": 0.0,
          "standard_len8_state_prefix_fraction": 0.0,
          "standard_len12_direct_accuracy": NaN,
          "standard_len12_direct_target_mass": NaN,
          "standard_len12_executor_accuracy": 0.0,
          "standard_len12_executor_target_mass": 0.010263217729516327,
          "standard_len12_init_accuracy": 0.0,
          "standard_len12_op_accuracy": 0.3515625,
          "standard_len12_arg_accuracy": 0.037760416666666664,
          "standard_len12_program_exact": 0.0,
          "standard_len12_state_accuracy": 0.005208333333333333,
          "standard_len12_state_all_exact": 0.0,
          "standard_len12_state_prefix_fraction": 0.0,
          "standard_len24_direct_accuracy": NaN,
          "standard_len24_direct_target_mass": NaN,
          "standard_len24_executor_accuracy": 0.015625,
          "standard_len24_executor_target_mass": 0.010263299336656928,
          "standard_len24_init_accuracy": 0.015625,
          "standard_len24_op_accuracy": 0.3424479166666667,
          "standard_len24_arg_accuracy": 0.026041666666666668,
          "standard_len24_program_exact": 0.0,
          "standard_len24_state_accuracy": 0.009114583333333334,
          "standard_len24_state_all_exact": 0.0,
          "standard_len24_state_prefix_fraction": 0.0006510416666666666,
          "paraphrase_len4_direct_accuracy": NaN,
          "paraphrase_len4_direct_target_mass": NaN,
          "paraphrase_len4_executor_accuracy": 0.03125,
          "paraphrase_len4_executor_target_mass": 0.01026039756834507,
          "paraphrase_len4_init_accuracy": 0.015625,
          "paraphrase_len4_op_accuracy": 0.31640625,
          "paraphrase_len4_arg_accuracy": 0.0390625,
          "paraphrase_len4_program_exact": 0.0,
          "paraphrase_len4_state_accuracy": 0.03125,
          "paraphrase_len4_state_all_exact": 0.0,
          "paraphrase_len4_state_prefix_fraction": 0.015625,
          "paraphrase_len8_direct_accuracy": NaN,
          "paraphrase_len8_direct_target_mass": NaN,
          "paraphrase_len8_executor_accuracy": 0.0,
          "paraphrase_len8_executor_target_mass": 0.010334869963116944,
          "paraphrase_len8_init_accuracy": 0.0,
          "paraphrase_len8_op_accuracy": 0.310546875,
          "paraphrase_len8_arg_accuracy": 0.03515625,
          "paraphrase_len8_program_exact": 0.0,
          "paraphrase_len8_state_accuracy": 0.01171875,
          "paraphrase_len8_state_all_exact": 0.0,
          "paraphrase_len8_state_prefix_fraction": 0.0,
          "paraphrase_len12_direct_accuracy": NaN,
          "paraphrase_len12_direct_target_mass": NaN,
          "paraphrase_len12_executor_accuracy": 0.0,
          "paraphrase_len12_executor_target_mass": 0.010482321842573583,
          "paraphrase_len12_init_accuracy": 0.015625,
          "paraphrase_len12_op_accuracy": 0.3580729166666667,
          "paraphrase_len12_arg_accuracy": 0.04296875,
          "paraphrase_len12_program_exact": 0.0,
          "paraphrase_len12_state_accuracy": 0.009114583333333334,
          "paraphrase_len12_state_all_exact": 0.0,
          "paraphrase_len12_state_prefix_fraction": 0.0,
          "paraphrase_len24_direct_accuracy": NaN,
          "paraphrase_len24_direct_target_mass": NaN,
          "paraphrase_len24_executor_accuracy": 0.0,
          "paraphrase_len24_executor_target_mass": 0.010264780139550567,
          "paraphrase_len24_init_accuracy": 0.0,
          "paraphrase_len24_op_accuracy": 0.3365885416666667,
          "paraphrase_len24_arg_accuracy": 0.037109375,
          "paraphrase_len24_program_exact": 0.0,
          "paraphrase_len24_state_accuracy": 0.0078125,
          "paraphrase_len24_state_all_exact": 0.0,
          "paraphrase_len24_state_prefix_fraction": 0.0,
          "paired_len4_direct_accuracy": NaN,
          "paired_len4_direct_target_mass": NaN,
          "paired_len4_executor_accuracy": 0.0,
          "paired_len4_executor_target_mass": 0.010261152929160744,
          "paired_len4_init_accuracy": 0.0,
          "paired_len4_op_accuracy": 0.28125,
          "paired_len4_arg_accuracy": 0.03515625,
          "paired_len4_program_exact": 0.0,
          "paired_len4_state_accuracy": 0.009765625,
          "paired_len4_state_all_exact": 0.0,
          "paired_len4_state_prefix_fraction": 0.001953125,
          "paired_len4_pair_true_answer_consistency": 1.0,
          "paired_len4_executor_pair_answer_consistency": 0.015625,
          "paired_len4_executor_pair_both_correct": 0.0,
          "paired_len4_compiler_pair_program_consistency": 0.015625,
          "paired_len4_compiler_pair_state_consistency": 0.015625,
          "paired_len8_direct_accuracy": NaN,
          "paired_len8_direct_target_mass": NaN,
          "paired_len8_executor_accuracy": 0.03125,
          "paired_len8_executor_target_mass": 0.010261766670737416,
          "paired_len8_init_accuracy": 0.03125,
          "paired_len8_op_accuracy": 0.333984375,
          "paired_len8_arg_accuracy": 0.0302734375,
          "paired_len8_program_exact": 0.0,
          "paired_len8_state_accuracy": 0.009765625,
          "paired_len8_state_all_exact": 0.0,
          "paired_len8_state_prefix_fraction": 0.0009765625,
          "paired_len8_pair_true_answer_consistency": 1.0,
          "paired_len8_executor_pair_answer_consistency": 0.015625,
          "paired_len8_executor_pair_both_correct": 0.0,
          "paired_len8_compiler_pair_program_consistency": 0.0,
          "paired_len8_compiler_pair_state_consistency": 0.0,
          "paired_len12_direct_accuracy": NaN,
          "paired_len12_direct_target_mass": NaN,
          "paired_len12_executor_accuracy": 0.0078125,
          "paired_len12_executor_target_mass": 0.010335352970287204,
          "paired_len12_init_accuracy": 0.0,
          "paired_len12_op_accuracy": 0.3268229166666667,
          "paired_len12_arg_accuracy": 0.0390625,
          "paired_len12_program_exact": 0.0,
          "paired_len12_state_accuracy": 0.014973958333333334,
          "paired_len12_state_all_exact": 0.0,
          "paired_len12_state_prefix_fraction": 0.0013020833333333333,
          "paired_len12_pair_true_answer_consistency": 1.0,
          "paired_len12_executor_pair_answer_consistency": 0.046875,
          "paired_len12_executor_pair_both_correct": 0.0,
          "paired_len12_compiler_pair_program_consistency": 0.0,
          "paired_len12_compiler_pair_state_consistency": 0.0,
          "paired_len24_direct_accuracy": NaN,
          "paired_len24_direct_target_mass": NaN,
          "paired_len24_executor_accuracy": 0.0234375,
          "paired_len24_executor_target_mass": 0.010333538462873548,
          "paired_len24_init_accuracy": 0.0,
          "paired_len24_op_accuracy": 0.3274739583333333,
          "paired_len24_arg_accuracy": 0.040690104166666664,
          "paired_len24_program_exact": 0.0,
          "paired_len24_state_accuracy": 0.009765625,
          "paired_len24_state_all_exact": 0.0,
          "paired_len24_state_prefix_fraction": 0.0,
          "paired_len24_pair_true_answer_consistency": 1.0,
          "paired_len24_executor_pair_answer_consistency": 0.0,
          "paired_len24_executor_pair_both_correct": 0.0,
          "paired_len24_compiler_pair_program_consistency": 0.0,
          "paired_len24_compiler_pair_state_consistency": 0.0
        },
        {
          "variant": "structured_trace_state_consistency",
          "stage": "structured_trace_state_consistency_short",
          "local_step": 80,
          "step": 80,
          "loss": 37.63404083251953,
          "trace_loss_active": true,
          "executor_loss_active": true,
          "state_loss_active": true,
          "lr_scale": 1.0,
          "executor_loss": 4.491498947143555,
          "state_loss": 4.4875264167785645,
          "state_train_accuracy": 0.0,
          "pair_program_loss": 0.012317043729126453,
          "pair_state_loss": 0.00013295430107973516,
          "init_loss": 4.0138840675354,
          "op_loss": 0.049567174166440964,
          "arg_loss": 3.1371536254882812,
          "standard_len4_direct_accuracy": NaN,
          "standard_len4_direct_target_mass": NaN,
          "standard_len4_executor_accuracy": 0.0,
          "standard_len4_executor_target_mass": 0.01030040392652154,
          "standard_len4_init_accuracy": 0.234375,
          "standard_len4_op_accuracy": 1.0,
          "standard_len4_arg_accuracy": 0.38671875,
          "standard_len4_program_exact": 0.0,
          "standard_len4_state_accuracy": 0.03515625,
          "standard_len4_state_all_exact": 0.0,
          "standard_len4_state_prefix_fraction": 0.03125,
          "standard_len8_direct_accuracy": NaN,
          "standard_len8_direct_target_mass": NaN,
          "standard_len8_executor_accuracy": 0.0,
          "standard_len8_executor_target_mass": 0.010298609733581543,
          "standard_len8_init_accuracy": 0.203125,
          "standard_len8_op_accuracy": 1.0,
          "standard_len8_arg_accuracy": 0.421875,
          "standard_len8_program_exact": 0.0,
          "standard_len8_state_accuracy": 0.02734375,
          "standard_len8_state_all_exact": 0.0,
          "standard_len8_state_prefix_fraction": 0.01953125,
          "standard_len12_direct_accuracy": NaN,
          "standard_len12_direct_target_mass": NaN,
          "standard_len12_executor_accuracy": 0.0,
          "standard_len12_executor_target_mass": 0.010282543837092817,
          "standard_len12_init_accuracy": 0.125,
          "standard_len12_op_accuracy": 0.9453125,
          "standard_len12_arg_accuracy": 0.3684895833333333,
          "standard_len12_program_exact": 0.0,
          "standard_len12_state_accuracy": 0.0234375,
          "standard_len12_state_all_exact": 0.0,
          "standard_len12_state_prefix_fraction": 0.015625,
          "standard_len24_direct_accuracy": NaN,
          "standard_len24_direct_target_mass": NaN,
          "standard_len24_executor_accuracy": 0.015625,
          "standard_len24_executor_target_mass": 0.010281694238074124,
          "standard_len24_init_accuracy": 0.140625,
          "standard_len24_op_accuracy": 0.8033854166666666,
          "standard_len24_arg_accuracy": 0.259765625,
          "standard_len24_program_exact": 0.0,
          "standard_len24_state_accuracy": 0.013671875,
          "standard_len24_state_all_exact": 0.0,
          "standard_len24_state_prefix_fraction": 0.001953125,
          "paraphrase_len4_direct_accuracy": NaN,
          "paraphrase_len4_direct_target_mass": NaN,
          "paraphrase_len4_executor_accuracy": 0.03125,
          "paraphrase_len4_executor_target_mass": 0.010285592288710177,
          "paraphrase_len4_init_accuracy": 0.21875,
          "paraphrase_len4_op_accuracy": 0.93359375,
          "paraphrase_len4_arg_accuracy": 0.3671875,
          "paraphrase_len4_program_exact": 0.0,
          "paraphrase_len4_state_accuracy": 0.05078125,
          "paraphrase_len4_state_all_exact": 0.0,
          "paraphrase_len4_state_prefix_fraction": 0.03125,
          "paraphrase_len8_direct_accuracy": NaN,
          "paraphrase_len8_direct_target_mass": NaN,
          "paraphrase_len8_executor_accuracy": 0.015625,
          "paraphrase_len8_executor_target_mass": 0.01030318159610033,
          "paraphrase_len8_init_accuracy": 0.125,
          "paraphrase_len8_op_accuracy": 0.822265625,
          "paraphrase_len8_arg_accuracy": 0.279296875,
          "paraphrase_len8_program_exact": 0.0,
          "paraphrase_len8_state_accuracy": 0.03125,
          "paraphrase_len8_state_all_exact": 0.0,
          "paraphrase_len8_state_prefix_fraction": 0.01953125,
          "paraphrase_len12_direct_accuracy": NaN,
          "paraphrase_len12_direct_target_mass": NaN,
          "paraphrase_len12_executor_accuracy": 0.015625,
          "paraphrase_len12_executor_target_mass": 0.010479909484274685,
          "paraphrase_len12_init_accuracy": 0.125,
          "paraphrase_len12_op_accuracy": 0.7604166666666666,
          "paraphrase_len12_arg_accuracy": 0.24739583333333334,
          "paraphrase_len12_program_exact": 0.0,
          "paraphrase_len12_state_accuracy": 0.016927083333333332,
          "paraphrase_len12_state_all_exact": 0.0,
          "paraphrase_len12_state_prefix_fraction": 0.00390625,
          "paraphrase_len24_direct_accuracy": NaN,
          "paraphrase_len24_direct_target_mass": NaN,
          "paraphrase_len24_executor_accuracy": 0.03125,
          "paraphrase_len24_executor_target_mass": 0.010277464287355542,
          "paraphrase_len24_init_accuracy": 0.203125,
          "paraphrase_len24_op_accuracy": 0.5592447916666666,
          "paraphrase_len24_arg_accuracy": 0.14713541666666666,
          "paraphrase_len24_program_exact": 0.0,
          "paraphrase_len24_state_accuracy": 0.012369791666666666,
          "paraphrase_len24_state_all_exact": 0.0,
          "paraphrase_len24_state_prefix_fraction": 0.007161458333333333,
          "paired_len4_direct_accuracy": NaN,
          "paired_len4_direct_target_mass": NaN,
          "paired_len4_executor_accuracy": 0.0,
          "paired_len4_executor_target_mass": 0.010294955922290683,
          "paired_len4_init_accuracy": 0.1015625,
          "paired_len4_op_accuracy": 0.97265625,
          "paired_len4_arg_accuracy": 0.3359375,
          "paired_len4_program_exact": 0.0,
          "paired_len4_state_accuracy": 0.005859375,
          "paired_len4_state_all_exact": 0.0,
          "paired_len4_state_prefix_fraction": 0.005859375,
          "paired_len4_pair_true_answer_consistency": 1.0,
          "paired_len4_executor_pair_answer_consistency": 0.46875,
          "paired_len4_executor_pair_both_correct": 0.0,
          "paired_len4_compiler_pair_program_consistency": 0.453125,
          "paired_len4_compiler_pair_state_consistency": 0.453125,
          "paired_len8_direct_accuracy": NaN,
          "paired_len8_direct_target_mass": NaN,
          "paired_len8_executor_accuracy": 0.015625,
          "paired_len8_executor_target_mass": 0.01028924627462402,
          "paired_len8_init_accuracy": 0.1953125,
          "paired_len8_op_accuracy": 0.9326171875,
          "paired_len8_arg_accuracy": 0.388671875,
          "paired_len8_program_exact": 0.0,
          "paired_len8_state_accuracy": 0.0283203125,
          "paired_len8_state_all_exact": 0.0,
          "paired_len8_state_prefix_fraction": 0.01953125,
          "paired_len8_pair_true_answer_consistency": 1.0,
          "paired_len8_executor_pair_answer_consistency": 0.0625,
          "paired_len8_executor_pair_both_correct": 0.0,
          "paired_len8_compiler_pair_program_consistency": 0.0625,
          "paired_len8_compiler_pair_state_consistency": 0.0625,
          "paired_len12_direct_accuracy": NaN,
          "paired_len12_direct_target_mass": NaN,
          "paired_len12_executor_accuracy": 0.015625,
          "paired_len12_executor_target_mass": 0.010420453734695911,
          "paired_len12_init_accuracy": 0.1953125,
          "paired_len12_op_accuracy": 0.853515625,
          "paired_len12_arg_accuracy": 0.3131510416666667,
          "paired_len12_program_exact": 0.0,
          "paired_len12_state_accuracy": 0.028645833333333332,
          "paired_len12_state_all_exact": 0.0,
          "paired_len12_state_prefix_fraction": 0.01953125,
          "paired_len12_pair_true_answer_consistency": 1.0,
          "paired_len12_executor_pair_answer_consistency": 0.015625,
          "paired_len12_executor_pair_both_correct": 0.0,
          "paired_len12_compiler_pair_program_consistency": 0.0,
          "paired_len12_compiler_pair_state_consistency": 0.0,
          "paired_len24_direct_accuracy": NaN,
          "paired_len24_direct_target_mass": NaN,
          "paired_len24_executor_accuracy": 0.0234375,
          "paired_len24_executor_target_mass": 0.01037386542884633,
          "paired_len24_init_accuracy": 0.1953125,
          "paired_len24_op_accuracy": 0.6634114583333334,
          "paired_len24_arg_accuracy": 0.19791666666666666,
          "paired_len24_program_exact": 0.0,
          "paired_len24_state_accuracy": 0.012369791666666666,
          "paired_len24_state_all_exact": 0.0,
          "paired_len24_state_prefix_fraction": 0.0022786458333333335,
          "paired_len24_pair_true_answer_consistency": 1.0,
          "paired_len24_executor_pair_answer_consistency": 0.015625,
          "paired_len24_executor_pair_both_correct": 0.0,
          "paired_len24_compiler_pair_program_consistency": 0.0,
          "paired_len24_compiler_pair_state_consistency": 0.0
        },
        {
          "variant": "structured_trace_state_consistency",
          "stage": "structured_trace_state_consistency_medium",
          "local_step": 1,
          "step": 81,
          "loss": 44.23200988769531,
          "trace_loss_active": true,
          "executor_loss_active": true,
          "state_loss_active": true,
          "lr_scale": 1.0,
          "executor_loss": 4.578611850738525,
          "state_loss": 4.575760364532471,
          "state_train_accuracy": 0.0,
          "pair_program_loss": 0.8843308687210083,
          "pair_state_loss": 8.893662015907466e-05,
          "init_loss": 5.105445384979248,
          "op_loss": 0.6174622774124146,
          "arg_loss": 3.4874799251556396,
          "standard_len4_direct_accuracy": NaN,
          "standard_len4_direct_target_mass": NaN,
          "standard_len4_executor_accuracy": 0.0,
          "standard_len4_executor_target_mass": 0.010300629306584597,
          "standard_len4_init_accuracy": 0.21875,
          "standard_len4_op_accuracy": 1.0,
          "standard_len4_arg_accuracy": 0.375,
          "standard_len4_program_exact": 0.0,
          "standard_len4_state_accuracy": 0.01953125,
          "standard_len4_state_all_exact": 0.0,
          "standard_len4_state_prefix_fraction": 0.015625,
          "standard_len8_direct_accuracy": NaN,
          "standard_len8_direct_target_mass": NaN,
          "standard_len8_executor_accuracy": 0.015625,
          "standard_len8_executor_target_mass": 0.010304971947334707,
          "standard_len8_init_accuracy": 0.21875,
          "standard_len8_op_accuracy": 1.0,
          "standard_len8_arg_accuracy": 0.412109375,
          "standard_len8_program_exact": 0.0,
          "standard_len8_state_accuracy": 0.033203125,
          "standard_len8_state_all_exact": 0.0,
          "standard_len8_state_prefix_fraction": 0.021484375,
          "standard_len12_direct_accuracy": NaN,
          "standard_len12_direct_target_mass": NaN,
          "standard_len12_executor_accuracy": 0.0,
          "standard_len12_executor_target_mass": 0.010281685623340309,
          "standard_len12_init_accuracy": 0.125,
          "standard_len12_op_accuracy": 0.9296875,
          "standard_len12_arg_accuracy": 0.3658854166666667,
          "standard_len12_program_exact": 0.0,
          "standard_len12_state_accuracy": 0.022135416666666668,
          "standard_len12_state_all_exact": 0.0,
          "standard_len12_state_prefix_fraction": 0.015625,
          "standard_len24_direct_accuracy": NaN,
          "standard_len24_direct_target_mass": NaN,
          "standard_len24_executor_accuracy": 0.015625,
          "standard_len24_executor_target_mass": 0.01028335781302303,
          "standard_len24_init_accuracy": 0.140625,
          "standard_len24_op_accuracy": 0.796875,
          "standard_len24_arg_accuracy": 0.2578125,
          "standard_len24_program_exact": 0.0,
          "standard_len24_state_accuracy": 0.013671875,
          "standard_len24_state_all_exact": 0.0,
          "standard_len24_state_prefix_fraction": 0.001953125,
          "paraphrase_len4_direct_accuracy": NaN,
          "paraphrase_len4_direct_target_mass": NaN,
          "paraphrase_len4_executor_accuracy": 0.03125,
          "paraphrase_len4_executor_target_mass": 0.010285517200827599,
          "paraphrase_len4_init_accuracy": 0.21875,
          "paraphrase_len4_op_accuracy": 0.93359375,
          "paraphrase_len4_arg_accuracy": 0.36328125,
          "paraphrase_len4_program_exact": 0.0,
          "paraphrase_len4_state_accuracy": 0.0546875,
          "paraphrase_len4_state_all_exact": 0.0,
          "paraphrase_len4_state_prefix_fraction": 0.02734375,
          "paraphrase_len8_direct_accuracy": NaN,
          "paraphrase_len8_direct_target_mass": NaN,
          "paraphrase_len8_executor_accuracy": 0.0,
          "paraphrase_len8_executor_target_mass": 0.010314087267033756,
          "paraphrase_len8_init_accuracy": 0.109375,
          "paraphrase_len8_op_accuracy": 0.806640625,
          "paraphrase_len8_arg_accuracy": 0.271484375,
          "paraphrase_len8_program_exact": 0.0,
          "paraphrase_len8_state_accuracy": 0.015625,
          "paraphrase_len8_state_all_exact": 0.0,
          "paraphrase_len8_state_prefix_fraction": 0.01171875,
          "paraphrase_len12_direct_accuracy": NaN,
          "paraphrase_len12_direct_target_mass": NaN,
          "paraphrase_len12_executor_accuracy": 0.0,
          "paraphrase_len12_executor_target_mass": 0.010485190316103399,
          "paraphrase_len12_init_accuracy": 0.15625,
          "paraphrase_len12_op_accuracy": 0.7473958333333334,
          "paraphrase_len12_arg_accuracy": 0.24609375,
          "paraphrase_len12_program_exact": 0.0,
          "paraphrase_len12_state_accuracy": 0.018229166666666668,
          "paraphrase_len12_state_all_exact": 0.0,
          "paraphrase_len12_state_prefix_fraction": 0.006510416666666667,
          "paraphrase_len24_direct_accuracy": NaN,
          "paraphrase_len24_direct_target_mass": NaN,
          "paraphrase_len24_executor_accuracy": 0.0,
          "paraphrase_len24_executor_target_mass": 0.010279944981448352,
          "paraphrase_len24_init_accuracy": 0.1875,
          "paraphrase_len24_op_accuracy": 0.5553385416666666,
          "paraphrase_len24_arg_accuracy": 0.14518229166666666,
          "paraphrase_len24_program_exact": 0.0,
          "paraphrase_len24_state_accuracy": 0.01171875,
          "paraphrase_len24_state_all_exact": 0.0,
          "paraphrase_len24_state_prefix_fraction": 0.007161458333333333,
          "paired_len4_direct_accuracy": NaN,
          "paired_len4_direct_target_mass": NaN,
          "paired_len4_executor_accuracy": 0.0,
          "paired_len4_executor_target_mass": 0.010296111984644085,
          "paired_len4_init_accuracy": 0.09375,
          "paired_len4_op_accuracy": 0.978515625,
          "paired_len4_arg_accuracy": 0.341796875,
          "paired_len4_program_exact": 0.0,
          "paired_len4_state_accuracy": 0.0078125,
          "paired_len4_state_all_exact": 0.0,
          "paired_len4_state_prefix_fraction": 0.0078125,
          "paired_len4_pair_true_answer_consistency": 1.0,
          "paired_len4_executor_pair_answer_consistency": 0.546875,
          "paired_len4_executor_pair_both_correct": 0.0,
          "paired_len4_compiler_pair_program_consistency": 0.53125,
          "paired_len4_compiler_pair_state_consistency": 0.546875,
          "paired_len8_direct_accuracy": NaN,
          "paired_len8_direct_target_mass": NaN,
          "paired_len8_executor_accuracy": 0.015625,
          "paired_len8_executor_target_mass": 0.010290001169778407,
          "paired_len8_init_accuracy": 0.203125,
          "paired_len8_op_accuracy": 0.9345703125,
          "paired_len8_arg_accuracy": 0.3876953125,
          "paired_len8_program_exact": 0.0,
          "paired_len8_state_accuracy": 0.029296875,
          "paired_len8_state_all_exact": 0.0,
          "paired_len8_state_prefix_fraction": 0.021484375,
          "paired_len8_pair_true_answer_consistency": 1.0,
          "paired_len8_executor_pair_answer_consistency": 0.078125,
          "paired_len8_executor_pair_both_correct": 0.0,
          "paired_len8_compiler_pair_program_consistency": 0.0625,
          "paired_len8_compiler_pair_state_consistency": 0.0625,
          "paired_len12_direct_accuracy": NaN,
          "paired_len12_direct_target_mass": NaN,
          "paired_len12_executor_accuracy": 0.0,
          "paired_len12_executor_target_mass": 0.010413670621346682,
          "paired_len12_init_accuracy": 0.1875,
          "paired_len12_op_accuracy": 0.841796875,
          "paired_len12_arg_accuracy": 0.3118489583333333,
          "paired_len12_program_exact": 0.0,
          "paired_len12_state_accuracy": 0.029947916666666668,
          "paired_len12_state_all_exact": 0.0,
          "paired_len12_state_prefix_fraction": 0.018880208333333332,
          "paired_len12_pair_true_answer_consistency": 1.0,
          "paired_len12_executor_pair_answer_consistency": 0.015625,
          "paired_len12_executor_pair_both_correct": 0.0,
          "paired_len12_compiler_pair_program_consistency": 0.015625,
          "paired_len12_compiler_pair_state_consistency": 0.015625,
          "paired_len24_direct_accuracy": NaN,
          "paired_len24_direct_target_mass": NaN,
          "paired_len24_executor_accuracy": 0.0078125,
          "paired_len24_executor_target_mass": 0.01035986456554383,
          "paired_len24_init_accuracy": 0.1796875,
          "paired_len24_op_accuracy": 0.658203125,
          "paired_len24_arg_accuracy": 0.2060546875,
          "paired_len24_program_exact": 0.0,
          "paired_len24_state_accuracy": 0.01171875,
          "paired_len24_state_all_exact": 0.0,
          "paired_len24_state_prefix_fraction": 0.0022786458333333335,
          "paired_len24_pair_true_answer_consistency": 1.0,
          "paired_len24_executor_pair_answer_consistency": 0.015625,
          "paired_len24_executor_pair_both_correct": 0.0,
          "paired_len24_compiler_pair_program_consistency": 0.0,
          "paired_len24_compiler_pair_state_consistency": 0.0
        },
        {
          "variant": "structured_trace_state_consistency",
          "stage": "structured_trace_state_consistency_medium",
          "local_step": 80,
          "step": 160,
          "loss": 10.033712387084961,
          "trace_loss_active": true,
          "executor_loss_active": true,
          "state_loss_active": true,
          "lr_scale": 1.0,
          "executor_loss": 2.654987096786499,
          "state_loss": 2.5366129875183105,
          "state_train_accuracy": 0.8333333134651184,
          "pair_program_loss": 0.0037166306283324957,
          "pair_state_loss": 0.00047525085392408073,
          "init_loss": 0.6143349409103394,
          "op_loss": 0.007244561333209276,
          "arg_loss": 0.5942298173904419,
          "standard_len4_direct_accuracy": NaN,
          "standard_len4_direct_target_mass": NaN,
          "standard_len4_executor_accuracy": 0.609375,
          "standard_len4_executor_target_mass": 0.023149261018261313,
          "standard_len4_init_accuracy": 0.671875,
          "standard_len4_op_accuracy": 1.0,
          "standard_len4_arg_accuracy": 0.95703125,
          "standard_len4_program_exact": 0.609375,
          "standard_len4_state_accuracy": 0.625,
          "standard_len4_state_all_exact": 0.609375,
          "standard_len4_state_prefix_fraction": 0.625,
          "standard_len8_direct_accuracy": NaN,
          "standard_len8_direct_target_mass": NaN,
          "standard_len8_executor_accuracy": 0.453125,
          "standard_len8_executor_target_mass": 0.011317304684780538,
          "standard_len8_init_accuracy": 0.578125,
          "standard_len8_op_accuracy": 1.0,
          "standard_len8_arg_accuracy": 0.95703125,
          "standard_len8_program_exact": 0.421875,
          "standard_len8_state_accuracy": 0.498046875,
          "standard_len8_state_all_exact": 0.421875,
          "standard_len8_state_prefix_fraction": 0.490234375,
          "standard_len12_direct_accuracy": NaN,
          "standard_len12_direct_target_mass": NaN,
          "standard_len12_executor_accuracy": 0.484375,
          "standard_len12_executor_target_mass": 0.010403058375231922,
          "standard_len12_init_accuracy": 0.6875,
          "standard_len12_op_accuracy": 1.0,
          "standard_len12_arg_accuracy": 0.9596354166666666,
          "standard_len12_program_exact": 0.484375,
          "standard_len12_state_accuracy": 0.5872395833333334,
          "standard_len12_state_all_exact": 0.484375,
          "standard_len12_state_prefix_fraction": 0.5872395833333334,
          "standard_len24_direct_accuracy": NaN,
          "standard_len24_direct_target_mass": NaN,
          "standard_len24_executor_accuracy": 0.0,
          "standard_len24_executor_target_mass": 0.010319901979528368,
          "standard_len24_init_accuracy": 0.625,
          "standard_len24_op_accuracy": 0.81640625,
          "standard_len24_arg_accuracy": 0.6569010416666666,
          "standard_len24_program_exact": 0.0,
          "standard_len24_state_accuracy": 0.24544270833333334,
          "standard_len24_state_all_exact": 0.0,
          "standard_len24_state_prefix_fraction": 0.24348958333333334,
          "paraphrase_len4_direct_accuracy": NaN,
          "paraphrase_len4_direct_target_mass": NaN,
          "paraphrase_len4_executor_accuracy": 0.515625,
          "paraphrase_len4_executor_target_mass": 0.02675232314504683,
          "paraphrase_len4_init_accuracy": 0.609375,
          "paraphrase_len4_op_accuracy": 1.0,
          "paraphrase_len4_arg_accuracy": 0.96875,
          "paraphrase_len4_program_exact": 0.515625,
          "paraphrase_len4_state_accuracy": 0.57421875,
          "paraphrase_len4_state_all_exact": 0.515625,
          "paraphrase_len4_state_prefix_fraction": 0.57421875,
          "paraphrase_len8_direct_accuracy": NaN,
          "paraphrase_len8_direct_target_mass": NaN,
          "paraphrase_len8_executor_accuracy": 0.234375,
          "paraphrase_len8_executor_target_mass": 0.010708913439884782,
          "paraphrase_len8_init_accuracy": 0.453125,
          "paraphrase_len8_op_accuracy": 0.96484375,
          "paraphrase_len8_arg_accuracy": 0.88671875,
          "paraphrase_len8_program_exact": 0.21875,
          "paraphrase_len8_state_accuracy": 0.330078125,
          "paraphrase_len8_state_all_exact": 0.21875,
          "paraphrase_len8_state_prefix_fraction": 0.318359375,
          "paraphrase_len12_direct_accuracy": NaN,
          "paraphrase_len12_direct_target_mass": NaN,
          "paraphrase_len12_executor_accuracy": 0.1875,
          "paraphrase_len12_executor_target_mass": 0.01036103954538703,
          "paraphrase_len12_init_accuracy": 0.5625,
          "paraphrase_len12_op_accuracy": 0.9609375,
          "paraphrase_len12_arg_accuracy": 0.8841145833333334,
          "paraphrase_len12_program_exact": 0.171875,
          "paraphrase_len12_state_accuracy": 0.3489583333333333,
          "paraphrase_len12_state_all_exact": 0.171875,
          "paraphrase_len12_state_prefix_fraction": 0.3463541666666667,
          "paraphrase_len24_direct_accuracy": NaN,
          "paraphrase_len24_direct_target_mass": NaN,
          "paraphrase_len24_executor_accuracy": 0.0,
          "paraphrase_len24_executor_target_mass": 0.010304185329005122,
          "paraphrase_len24_init_accuracy": 0.546875,
          "paraphrase_len24_op_accuracy": 0.701171875,
          "paraphrase_len24_arg_accuracy": 0.529296875,
          "paraphrase_len24_program_exact": 0.0,
          "paraphrase_len24_state_accuracy": 0.21809895833333334,
          "paraphrase_len24_state_all_exact": 0.0,
          "paraphrase_len24_state_prefix_fraction": 0.21028645833333334,
          "paired_len4_direct_accuracy": NaN,
          "paired_len4_direct_target_mass": NaN,
          "paired_len4_executor_accuracy": 0.5,
          "paired_len4_executor_target_mass": 0.01750925072701648,
          "paired_len4_init_accuracy": 0.6015625,
          "paired_len4_op_accuracy": 1.0,
          "paired_len4_arg_accuracy": 0.951171875,
          "paired_len4_program_exact": 0.5,
          "paired_len4_state_accuracy": 0.537109375,
          "paired_len4_state_all_exact": 0.5,
          "paired_len4_state_prefix_fraction": 0.537109375,
          "paired_len4_pair_true_answer_consistency": 1.0,
          "paired_len4_executor_pair_answer_consistency": 0.90625,
          "paired_len4_executor_pair_both_correct": 0.46875,
          "paired_len4_compiler_pair_program_consistency": 0.90625,
          "paired_len4_compiler_pair_state_consistency": 0.90625,
          "paired_len8_direct_accuracy": NaN,
          "paired_len8_direct_target_mass": NaN,
          "paired_len8_executor_accuracy": 0.4921875,
          "paired_len8_executor_target_mass": 0.012341370049398392,
          "paired_len8_init_accuracy": 0.6328125,
          "paired_len8_op_accuracy": 0.9912109375,
          "paired_len8_arg_accuracy": 0.94921875,
          "paired_len8_program_exact": 0.4921875,
          "paired_len8_state_accuracy": 0.552734375,
          "paired_len8_state_all_exact": 0.4921875,
          "paired_len8_state_prefix_fraction": 0.552734375,
          "paired_len8_pair_true_answer_consistency": 1.0,
          "paired_len8_executor_pair_answer_consistency": 0.875,
          "paired_len8_executor_pair_both_correct": 0.46875,
          "paired_len8_compiler_pair_program_consistency": 0.875,
          "paired_len8_compiler_pair_state_consistency": 0.875,
          "paired_len12_direct_accuracy": NaN,
          "paired_len12_direct_target_mass": NaN,
          "paired_len12_executor_accuracy": 0.328125,
          "paired_len12_executor_target_mass": 0.010525063204113394,
          "paired_len12_init_accuracy": 0.7265625,
          "paired_len12_op_accuracy": 0.9694010416666666,
          "paired_len12_arg_accuracy": 0.8932291666666666,
          "paired_len12_program_exact": 0.328125,
          "paired_len12_state_accuracy": 0.4973958333333333,
          "paired_len12_state_all_exact": 0.328125,
          "paired_len12_state_prefix_fraction": 0.49609375,
          "paired_len12_pair_true_answer_consistency": 1.0,
          "paired_len12_executor_pair_answer_consistency": 0.5625,
          "paired_len12_executor_pair_both_correct": 0.296875,
          "paired_len12_compiler_pair_program_consistency": 0.5625,
          "paired_len12_compiler_pair_state_consistency": 0.5625,
          "paired_len24_direct_accuracy": NaN,
          "paired_len24_direct_target_mass": NaN,
          "paired_len24_executor_accuracy": 0.0078125,
          "paired_len24_executor_target_mass": 0.010303019313141704,
          "paired_len24_init_accuracy": 0.5078125,
          "paired_len24_op_accuracy": 0.7412109375,
          "paired_len24_arg_accuracy": 0.5875651041666666,
          "paired_len24_program_exact": 0.0,
          "paired_len24_state_accuracy": 0.22819010416666666,
          "paired_len24_state_all_exact": 0.0,
          "paired_len24_state_prefix_fraction": 0.2216796875,
          "paired_len24_pair_true_answer_consistency": 1.0,
          "paired_len24_executor_pair_answer_consistency": 0.0,
          "paired_len24_executor_pair_both_correct": 0.0,
          "paired_len24_compiler_pair_program_consistency": 0.0,
          "paired_len24_compiler_pair_state_consistency": 0.0
        },
        {
          "variant": "structured_trace_state_consistency",
          "stage": "structured_trace_state_consistency_train",
          "local_step": 1,
          "step": 161,
          "loss": 27.609663009643555,
          "trace_loss_active": true,
          "executor_loss_active": true,
          "state_loss_active": true,
          "lr_scale": 1.0,
          "executor_loss": 4.575992107391357,
          "state_loss": 4.440532207489014,
          "state_train_accuracy": 0.23529411852359772,
          "pair_program_loss": 0.015211493708193302,
          "pair_state_loss": 6.0324236983433366e-05,
          "init_loss": 3.531378746032715,
          "op_loss": 0.01474193949252367,
          "arg_loss": 1.11283278465271,
          "standard_len4_direct_accuracy": NaN,
          "standard_len4_direct_target_mass": NaN,
          "standard_len4_executor_accuracy": 0.640625,
          "standard_len4_executor_target_mass": 0.024013491114601493,
          "standard_len4_init_accuracy": 0.671875,
          "standard_len4_op_accuracy": 1.0,
          "standard_len4_arg_accuracy": 0.98046875,
          "standard_len4_program_exact": 0.640625,
          "standard_len4_state_accuracy": 0.65625,
          "standard_len4_state_all_exact": 0.640625,
          "standard_len4_state_prefix_fraction": 0.65625,
          "standard_len8_direct_accuracy": NaN,
          "standard_len8_direct_target_mass": NaN,
          "standard_len8_executor_accuracy": 0.4375,
          "standard_len8_executor_target_mass": 0.011457669665105641,
          "standard_len8_init_accuracy": 0.546875,
          "standard_len8_op_accuracy": 1.0,
          "standard_len8_arg_accuracy": 0.966796875,
          "standard_len8_program_exact": 0.421875,
          "standard_len8_state_accuracy": 0.4765625,
          "standard_len8_state_all_exact": 0.421875,
          "standard_len8_state_prefix_fraction": 0.474609375,
          "standard_len12_direct_accuracy": NaN,
          "standard_len12_direct_target_mass": NaN,
          "standard_len12_executor_accuracy": 0.53125,
          "standard_len12_executor_target_mass": 0.01041950355283916,
          "standard_len12_init_accuracy": 0.671875,
          "standard_len12_op_accuracy": 1.0,
          "standard_len12_arg_accuracy": 0.97265625,
          "standard_len12_program_exact": 0.53125,
          "standard_len12_state_accuracy": 0.6119791666666666,
          "standard_len12_state_all_exact": 0.53125,
          "standard_len12_state_prefix_fraction": 0.6119791666666666,
          "standard_len24_direct_accuracy": NaN,
          "standard_len24_direct_target_mass": NaN,
          "standard_len24_executor_accuracy": 0.0,
          "standard_len24_executor_target_mass": 0.010321857291273773,
          "standard_len24_init_accuracy": 0.625,
          "standard_len24_op_accuracy": 0.81640625,
          "standard_len24_arg_accuracy": 0.6640625,
          "standard_len24_program_exact": 0.0,
          "standard_len24_state_accuracy": 0.2805989583333333,
          "standard_len24_state_all_exact": 0.0,
          "standard_len24_state_prefix_fraction": 0.2779947916666667,
          "paraphrase_len4_direct_accuracy": NaN,
          "paraphrase_len4_direct_target_mass": NaN,
          "paraphrase_len4_executor_accuracy": 0.5625,
          "paraphrase_len4_executor_target_mass": 0.027572402032092214,
          "paraphrase_len4_init_accuracy": 0.640625,
          "paraphrase_len4_op_accuracy": 1.0,
          "paraphrase_len4_arg_accuracy": 0.97265625,
          "paraphrase_len4_program_exact": 0.5625,
          "paraphrase_len4_state_accuracy": 0.609375,
          "paraphrase_len4_state_all_exact": 0.5625,
          "paraphrase_len4_state_prefix_fraction": 0.609375,
          "paraphrase_len8_direct_accuracy": NaN,
          "paraphrase_len8_direct_target_mass": NaN,
          "paraphrase_len8_executor_accuracy": 0.234375,
          "paraphrase_len8_executor_target_mass": 0.010779290460050106,
          "paraphrase_len8_init_accuracy": 0.453125,
          "paraphrase_len8_op_accuracy": 0.962890625,
          "paraphrase_len8_arg_accuracy": 0.90234375,
          "paraphrase_len8_program_exact": 0.234375,
          "paraphrase_len8_state_accuracy": 0.330078125,
          "paraphrase_len8_state_all_exact": 0.234375,
          "paraphrase_len8_state_prefix_fraction": 0.330078125,
          "paraphrase_len12_direct_accuracy": NaN,
          "paraphrase_len12_direct_target_mass": NaN,
          "paraphrase_len12_executor_accuracy": 0.265625,
          "paraphrase_len12_executor_target_mass": 0.01037182321306318,
          "paraphrase_len12_init_accuracy": 0.5625,
          "paraphrase_len12_op_accuracy": 0.9609375,
          "paraphrase_len12_arg_accuracy": 0.8971354166666666,
          "paraphrase_len12_program_exact": 0.25,
          "paraphrase_len12_state_accuracy": 0.3841145833333333,
          "paraphrase_len12_state_all_exact": 0.25,
          "paraphrase_len12_state_prefix_fraction": 0.3828125,
          "paraphrase_len24_direct_accuracy": NaN,
          "paraphrase_len24_direct_target_mass": NaN,
          "paraphrase_len24_executor_accuracy": 0.0,
          "paraphrase_len24_executor_target_mass": 0.010304395109415054,
          "paraphrase_len24_init_accuracy": 0.53125,
          "paraphrase_len24_op_accuracy": 0.7024739583333334,
          "paraphrase_len24_arg_accuracy": 0.537109375,
          "paraphrase_len24_program_exact": 0.0,
          "paraphrase_len24_state_accuracy": 0.21549479166666666,
          "paraphrase_len24_state_all_exact": 0.0,
          "paraphrase_len24_state_prefix_fraction": 0.21028645833333334,
          "paired_len4_direct_accuracy": NaN,
          "paired_len4_direct_target_mass": NaN,
          "paired_len4_executor_accuracy": 0.4921875,
          "paired_len4_executor_target_mass": 0.018111868179403245,
          "paired_len4_init_accuracy": 0.578125,
          "paired_len4_op_accuracy": 1.0,
          "paired_len4_arg_accuracy": 0.958984375,
          "paired_len4_program_exact": 0.4921875,
          "paired_len4_state_accuracy": 0.525390625,
          "paired_len4_state_all_exact": 0.4921875,
          "paired_len4_state_prefix_fraction": 0.525390625,
          "paired_len4_pair_true_answer_consistency": 1.0,
          "paired_len4_executor_pair_answer_consistency": 0.9375,
          "paired_len4_executor_pair_both_correct": 0.46875,
          "paired_len4_compiler_pair_program_consistency": 0.9375,
          "paired_len4_compiler_pair_state_consistency": 0.9375,
          "paired_len8_direct_accuracy": NaN,
          "paired_len8_direct_target_mass": NaN,
          "paired_len8_executor_accuracy": 0.53125,
          "paired_len8_executor_target_mass": 0.012587831530254334,
          "paired_len8_init_accuracy": 0.6328125,
          "paired_len8_op_accuracy": 0.9931640625,
          "paired_len8_arg_accuracy": 0.958984375,
          "paired_len8_program_exact": 0.53125,
          "paired_len8_state_accuracy": 0.580078125,
          "paired_len8_state_all_exact": 0.53125,
          "paired_len8_state_prefix_fraction": 0.580078125,
          "paired_len8_pair_true_answer_consistency": 1.0,
          "paired_len8_executor_pair_answer_consistency": 0.859375,
          "paired_len8_executor_pair_both_correct": 0.515625,
          "paired_len8_compiler_pair_program_consistency": 0.859375,
          "paired_len8_compiler_pair_state_consistency": 0.859375,
          "paired_len12_direct_accuracy": NaN,
          "paired_len12_direct_target_mass": NaN,
          "paired_len12_executor_accuracy": 0.4140625,
          "paired_len12_executor_target_mass": 0.010555185028351843,
          "paired_len12_init_accuracy": 0.7265625,
          "paired_len12_op_accuracy": 0.97265625,
          "paired_len12_arg_accuracy": 0.9140625,
          "paired_len12_program_exact": 0.40625,
          "paired_len12_state_accuracy": 0.5703125,
          "paired_len12_state_all_exact": 0.40625,
          "paired_len12_state_prefix_fraction": 0.5690104166666666,
          "paired_len12_pair_true_answer_consistency": 1.0,
          "paired_len12_executor_pair_answer_consistency": 0.578125,
          "paired_len12_executor_pair_both_correct": 0.359375,
          "paired_len12_compiler_pair_program_consistency": 0.578125,
          "paired_len12_compiler_pair_state_consistency": 0.578125,
          "paired_len24_direct_accuracy": NaN,
          "paired_len24_direct_target_mass": NaN,
          "paired_len24_executor_accuracy": 0.0078125,
          "paired_len24_executor_target_mass": 0.010302758193574846,
          "paired_len24_init_accuracy": 0.484375,
          "paired_len24_op_accuracy": 0.740234375,
          "paired_len24_arg_accuracy": 0.5930989583333334,
          "paired_len24_program_exact": 0.0,
          "paired_len24_state_accuracy": 0.21256510416666666,
          "paired_len24_state_all_exact": 0.0,
          "paired_len24_state_prefix_fraction": 0.20475260416666666,
          "paired_len24_pair_true_answer_consistency": 1.0,
          "paired_len24_executor_pair_answer_consistency": 0.0,
          "paired_len24_executor_pair_both_correct": 0.0,
          "paired_len24_compiler_pair_program_consistency": 0.0,
          "paired_len24_compiler_pair_state_consistency": 0.0
        },
        {
          "variant": "structured_trace_state_consistency",
          "stage": "structured_trace_state_consistency_train",
          "local_step": 80,
          "step": 240,
          "loss": 7.763700485229492,
          "trace_loss_active": true,
          "executor_loss_active": true,
          "state_loss_active": true,
          "lr_scale": 1.0,
          "executor_loss": 1.5169703960418701,
          "state_loss": 1.3111953735351562,
          "state_train_accuracy": 1.0,
          "pair_program_loss": 0.0011123993899673223,
          "pair_state_loss": 0.00047938033821992576,
          "init_loss": 1.1815276145935059,
          "op_loss": 0.0011618572752922773,
          "arg_loss": 0.051977794617414474,
          "standard_len4_direct_accuracy": NaN,
          "standard_len4_direct_target_mass": NaN,
          "standard_len4_executor_accuracy": 0.859375,
          "standard_len4_executor_target_mass": 0.2408065591007471,
          "standard_len4_init_accuracy": 0.859375,
          "standard_len4_op_accuracy": 1.0,
          "standard_len4_arg_accuracy": 1.0,
          "standard_len4_program_exact": 0.859375,
          "standard_len4_state_accuracy": 0.859375,
          "standard_len4_state_all_exact": 0.859375,
          "standard_len4_state_prefix_fraction": 0.859375,
          "standard_len8_direct_accuracy": NaN,
          "standard_len8_direct_target_mass": NaN,
          "standard_len8_executor_accuracy": 0.796875,
          "standard_len8_executor_target_mass": 0.1302952878177166,
          "standard_len8_init_accuracy": 0.796875,
          "standard_len8_op_accuracy": 1.0,
          "standard_len8_arg_accuracy": 1.0,
          "standard_len8_program_exact": 0.796875,
          "standard_len8_state_accuracy": 0.796875,
          "standard_len8_state_all_exact": 0.796875,
          "standard_len8_state_prefix_fraction": 0.796875,
          "standard_len12_direct_accuracy": NaN,
          "standard_len12_direct_target_mass": NaN,
          "standard_len12_executor_accuracy": 0.859375,
          "standard_len12_executor_target_mass": 0.10725165531039238,
          "standard_len12_init_accuracy": 0.859375,
          "standard_len12_op_accuracy": 1.0,
          "standard_len12_arg_accuracy": 1.0,
          "standard_len12_program_exact": 0.859375,
          "standard_len12_state_accuracy": 0.859375,
          "standard_len12_state_all_exact": 0.859375,
          "standard_len12_state_prefix_fraction": 0.859375,
          "standard_len24_direct_accuracy": NaN,
          "standard_len24_direct_target_mass": NaN,
          "standard_len24_executor_accuracy": 0.078125,
          "standard_len24_executor_target_mass": 0.011610763031058013,
          "standard_len24_init_accuracy": 0.84375,
          "standard_len24_op_accuracy": 0.9127604166666666,
          "standard_len24_arg_accuracy": 0.85546875,
          "standard_len24_program_exact": 0.0625,
          "standard_len24_state_accuracy": 0.6432291666666666,
          "standard_len24_state_all_exact": 0.0625,
          "standard_len24_state_prefix_fraction": 0.6412760416666666,
          "paraphrase_len4_direct_accuracy": NaN,
          "paraphrase_len4_direct_target_mass": NaN,
          "paraphrase_len4_executor_accuracy": 0.828125,
          "paraphrase_len4_executor_target_mass": 0.21950156427919865,
          "paraphrase_len4_init_accuracy": 0.828125,
          "paraphrase_len4_op_accuracy": 1.0,
          "paraphrase_len4_arg_accuracy": 1.0,
          "paraphrase_len4_program_exact": 0.828125,
          "paraphrase_len4_state_accuracy": 0.828125,
          "paraphrase_len4_state_all_exact": 0.828125,
          "paraphrase_len4_state_prefix_fraction": 0.828125,
          "paraphrase_len8_direct_accuracy": NaN,
          "paraphrase_len8_direct_target_mass": NaN,
          "paraphrase_len8_executor_accuracy": 0.75,
          "paraphrase_len8_executor_target_mass": 0.12286304496228695,
          "paraphrase_len8_init_accuracy": 0.75,
          "paraphrase_len8_op_accuracy": 1.0,
          "paraphrase_len8_arg_accuracy": 1.0,
          "paraphrase_len8_program_exact": 0.75,
          "paraphrase_len8_state_accuracy": 0.75,
          "paraphrase_len8_state_all_exact": 0.75,
          "paraphrase_len8_state_prefix_fraction": 0.75,
          "paraphrase_len12_direct_accuracy": NaN,
          "paraphrase_len12_direct_target_mass": NaN,
          "paraphrase_len12_executor_accuracy": 0.78125,
          "paraphrase_len12_executor_target_mass": 0.10984228271991014,
          "paraphrase_len12_init_accuracy": 0.78125,
          "paraphrase_len12_op_accuracy": 1.0,
          "paraphrase_len12_arg_accuracy": 1.0,
          "paraphrase_len12_program_exact": 0.78125,
          "paraphrase_len12_state_accuracy": 0.78125,
          "paraphrase_len12_state_all_exact": 0.78125,
          "paraphrase_len12_state_prefix_fraction": 0.78125,
          "paraphrase_len24_direct_accuracy": NaN,
          "paraphrase_len24_direct_target_mass": NaN,
          "paraphrase_len24_executor_accuracy": 0.09375,
          "paraphrase_len24_executor_target_mass": 0.01400190859567374,
          "paraphrase_len24_init_accuracy": 0.703125,
          "paraphrase_len24_op_accuracy": 0.892578125,
          "paraphrase_len24_arg_accuracy": 0.8294270833333334,
          "paraphrase_len24_program_exact": 0.078125,
          "paraphrase_len24_state_accuracy": 0.5221354166666666,
          "paraphrase_len24_state_all_exact": 0.078125,
          "paraphrase_len24_state_prefix_fraction": 0.5188802083333334,
          "paired_len4_direct_accuracy": NaN,
          "paired_len4_direct_target_mass": NaN,
          "paired_len4_executor_accuracy": 0.765625,
          "paired_len4_executor_target_mass": 0.15648487000726163,
          "paired_len4_init_accuracy": 0.765625,
          "paired_len4_op_accuracy": 1.0,
          "paired_len4_arg_accuracy": 1.0,
          "paired_len4_program_exact": 0.765625,
          "paired_len4_state_accuracy": 0.765625,
          "paired_len4_state_all_exact": 0.765625,
          "paired_len4_state_prefix_fraction": 0.765625,
          "paired_len4_pair_true_answer_consistency": 1.0,
          "paired_len4_executor_pair_answer_consistency": 0.96875,
          "paired_len4_executor_pair_both_correct": 0.765625,
          "paired_len4_compiler_pair_program_consistency": 0.96875,
          "paired_len4_compiler_pair_state_consistency": 0.96875,
          "paired_len8_direct_accuracy": NaN,
          "paired_len8_direct_target_mass": NaN,
          "paired_len8_executor_accuracy": 0.875,
          "paired_len8_executor_target_mass": 0.18304786505177617,
          "paired_len8_init_accuracy": 0.875,
          "paired_len8_op_accuracy": 1.0,
          "paired_len8_arg_accuracy": 1.0,
          "paired_len8_program_exact": 0.875,
          "paired_len8_state_accuracy": 0.875,
          "paired_len8_state_all_exact": 0.875,
          "paired_len8_state_prefix_fraction": 0.875,
          "paired_len8_pair_true_answer_consistency": 1.0,
          "paired_len8_executor_pair_answer_consistency": 1.0,
          "paired_len8_executor_pair_both_correct": 0.875,
          "paired_len8_compiler_pair_program_consistency": 1.0,
          "paired_len8_compiler_pair_state_consistency": 1.0,
          "paired_len12_direct_accuracy": NaN,
          "paired_len12_direct_target_mass": NaN,
          "paired_len12_executor_accuracy": 0.8203125,
          "paired_len12_executor_target_mass": 0.14072793535888195,
          "paired_len12_init_accuracy": 0.84375,
          "paired_len12_op_accuracy": 0.9993489583333334,
          "paired_len12_arg_accuracy": 0.9986979166666666,
          "paired_len12_program_exact": 0.8203125,
          "paired_len12_state_accuracy": 0.83984375,
          "paired_len12_state_all_exact": 0.8203125,
          "paired_len12_state_prefix_fraction": 0.83984375,
          "paired_len12_pair_true_answer_consistency": 1.0,
          "paired_len12_executor_pair_answer_consistency": 0.953125,
          "paired_len12_executor_pair_both_correct": 0.796875,
          "paired_len12_compiler_pair_program_consistency": 0.953125,
          "paired_len12_compiler_pair_state_consistency": 0.953125,
          "paired_len24_direct_accuracy": NaN,
          "paired_len24_direct_target_mass": NaN,
          "paired_len24_executor_accuracy": 0.09375,
          "paired_len24_executor_target_mass": 0.012367482180707157,
          "paired_len24_init_accuracy": 0.890625,
          "paired_len24_op_accuracy": 0.9010416666666666,
          "paired_len24_arg_accuracy": 0.8365885416666666,
          "paired_len24_program_exact": 0.078125,
          "paired_len24_state_accuracy": 0.6630859375,
          "paired_len24_state_all_exact": 0.078125,
          "paired_len24_state_prefix_fraction": 0.6572265625,
          "paired_len24_pair_true_answer_consistency": 1.0,
          "paired_len24_executor_pair_answer_consistency": 0.03125,
          "paired_len24_executor_pair_both_correct": 0.03125,
          "paired_len24_compiler_pair_program_consistency": 0.03125,
          "paired_len24_compiler_pair_state_consistency": 0.03125
        }
      ],
      "final_metrics": {
        "standard_len4": {
          "variant": "structured_trace_state_consistency",
          "n": 64.0,
          "direct_accuracy": NaN,
          "direct_target_mass": NaN,
          "executor_accuracy": 0.859375,
          "executor_target_mass": 0.2408065591007471,
          "init_accuracy": 0.859375,
          "op_accuracy": 1.0,
          "arg_accuracy": 1.0,
          "program_exact": 0.859375,
          "state_accuracy": 0.859375,
          "state_all_exact": 0.859375,
          "state_prefix_fraction": 0.859375
        },
        "standard_len8": {
          "variant": "structured_trace_state_consistency",
          "n": 64.0,
          "direct_accuracy": NaN,
          "direct_target_mass": NaN,
          "executor_accuracy": 0.796875,
          "executor_target_mass": 0.1302952878177166,
          "init_accuracy": 0.796875,
          "op_accuracy": 1.0,
          "arg_accuracy": 1.0,
          "program_exact": 0.796875,
          "state_accuracy": 0.796875,
          "state_all_exact": 0.796875,
          "state_prefix_fraction": 0.796875
        },
        "standard_len12": {
          "variant": "structured_trace_state_consistency",
          "n": 64.0,
          "direct_accuracy": NaN,
          "direct_target_mass": NaN,
          "executor_accuracy": 0.859375,
          "executor_target_mass": 0.10725165531039238,
          "init_accuracy": 0.859375,
          "op_accuracy": 1.0,
          "arg_accuracy": 1.0,
          "program_exact": 0.859375,
          "state_accuracy": 0.859375,
          "state_all_exact": 0.859375,
          "state_prefix_fraction": 0.859375
        },
        "standard_len24": {
          "variant": "structured_trace_state_consistency",
          "n": 64.0,
          "direct_accuracy": NaN,
          "direct_target_mass": NaN,
          "executor_accuracy": 0.078125,
          "executor_target_mass": 0.011610763031058013,
          "init_accuracy": 0.84375,
          "op_accuracy": 0.9127604166666666,
          "arg_accuracy": 0.85546875,
          "program_exact": 0.0625,
          "state_accuracy": 0.6432291666666666,
          "state_all_exact": 0.0625,
          "state_prefix_fraction": 0.6412760416666666
        },
        "paraphrase_len4": {
          "variant": "structured_trace_state_consistency",
          "n": 64.0,
          "direct_accuracy": NaN,
          "direct_target_mass": NaN,
          "executor_accuracy": 0.828125,
          "executor_target_mass": 0.21950156427919865,
          "init_accuracy": 0.828125,
          "op_accuracy": 1.0,
          "arg_accuracy": 1.0,
          "program_exact": 0.828125,
          "state_accuracy": 0.828125,
          "state_all_exact": 0.828125,
          "state_prefix_fraction": 0.828125
        },
        "paraphrase_len8": {
          "variant": "structured_trace_state_consistency",
          "n": 64.0,
          "direct_accuracy": NaN,
          "direct_target_mass": NaN,
          "executor_accuracy": 0.75,
          "executor_target_mass": 0.12286304496228695,
          "init_accuracy": 0.75,
          "op_accuracy": 1.0,
          "arg_accuracy": 1.0,
          "program_exact": 0.75,
          "state_accuracy": 0.75,
          "state_all_exact": 0.75,
          "state_prefix_fraction": 0.75
        },
        "paraphrase_len12": {
          "variant": "structured_trace_state_consistency",
          "n": 64.0,
          "direct_accuracy": NaN,
          "direct_target_mass": NaN,
          "executor_accuracy": 0.78125,
          "executor_target_mass": 0.10984228271991014,
          "init_accuracy": 0.78125,
          "op_accuracy": 1.0,
          "arg_accuracy": 1.0,
          "program_exact": 0.78125,
          "state_accuracy": 0.78125,
          "state_all_exact": 0.78125,
          "state_prefix_fraction": 0.78125
        },
        "paraphrase_len24": {
          "variant": "structured_trace_state_consistency",
          "n": 64.0,
          "direct_accuracy": NaN,
          "direct_target_mass": NaN,
          "executor_accuracy": 0.09375,
          "executor_target_mass": 0.01400190859567374,
          "init_accuracy": 0.703125,
          "op_accuracy": 0.892578125,
          "arg_accuracy": 0.8294270833333334,
          "program_exact": 0.078125,
          "state_accuracy": 0.5221354166666666,
          "state_all_exact": 0.078125,
          "state_prefix_fraction": 0.5188802083333334
        },
        "paired_len4": {
          "variant": "structured_trace_state_consistency",
          "n": 128.0,
          "direct_accuracy": NaN,
          "direct_target_mass": NaN,
          "executor_accuracy": 0.765625,
          "executor_target_mass": 0.15648487000726163,
          "init_accuracy": 0.765625,
          "op_accuracy": 1.0,
          "arg_accuracy": 1.0,
          "program_exact": 0.765625,
          "state_accuracy": 0.765625,
          "state_all_exact": 0.765625,
          "state_prefix_fraction": 0.765625,
          "pair_true_answer_consistency": 1.0,
          "executor_pair_answer_consistency": 0.96875,
          "executor_pair_both_correct": 0.765625,
          "compiler_pair_program_consistency": 0.96875,
          "compiler_pair_state_consistency": 0.96875
        },
        "paired_len8": {
          "variant": "structured_trace_state_consistency",
          "n": 128.0,
          "direct_accuracy": NaN,
          "direct_target_mass": NaN,
          "executor_accuracy": 0.875,
          "executor_target_mass": 0.18304786505177617,
          "init_accuracy": 0.875,
          "op_accuracy": 1.0,
          "arg_accuracy": 1.0,
          "program_exact": 0.875,
          "state_accuracy": 0.875,
          "state_all_exact": 0.875,
          "state_prefix_fraction": 0.875,
          "pair_true_answer_consistency": 1.0,
          "executor_pair_answer_consistency": 1.0,
          "executor_pair_both_correct": 0.875,
          "compiler_pair_program_consistency": 1.0,
          "compiler_pair_state_consistency": 1.0
        },
        "paired_len12": {
          "variant": "structured_trace_state_consistency",
          "n": 128.0,
          "direct_accuracy": NaN,
          "direct_target_mass": NaN,
          "executor_accuracy": 0.8203125,
          "executor_target_mass": 0.14072793535888195,
          "init_accuracy": 0.84375,
          "op_accuracy": 0.9993489583333334,
          "arg_accuracy": 0.9986979166666666,
          "program_exact": 0.8203125,
          "state_accuracy": 0.83984375,
          "state_all_exact": 0.8203125,
          "state_prefix_fraction": 0.83984375,
          "pair_true_answer_consistency": 1.0,
          "executor_pair_answer_consistency": 0.953125,
          "executor_pair_both_correct": 0.796875,
          "compiler_pair_program_consistency": 0.953125,
          "compiler_pair_state_consistency": 0.953125
        },
        "paired_len24": {
          "variant": "structured_trace_state_consistency",
          "n": 128.0,
          "direct_accuracy": NaN,
          "direct_target_mass": NaN,
          "executor_accuracy": 0.09375,
          "executor_target_mass": 0.012367482180707157,
          "init_accuracy": 0.890625,
          "op_accuracy": 0.9010416666666666,
          "arg_accuracy": 0.8365885416666666,
          "program_exact": 0.078125,
          "state_accuracy": 0.6630859375,
          "state_all_exact": 0.078125,
          "state_prefix_fraction": 0.6572265625,
          "pair_true_answer_consistency": 1.0,
          "executor_pair_answer_consistency": 0.03125,
          "executor_pair_both_correct": 0.03125,
          "compiler_pair_program_consistency": 0.03125,
          "compiler_pair_state_consistency": 0.03125
        }
      }
    }
  }
}