{
  "args": {
    "model_id": "Qwen/Qwen3.5-4B",
    "text_model_fallback": "Qwen/Qwen3-4B",
    "model_class": "auto",
    "load_in_4bit": true,
    "torch_dtype": "bf16",
    "device_map": "auto",
    "modulus": 97,
    "max_steps": 24,
    "train_min_len": 1,
    "train_max_len": 12,
    "answer_train_min_len": 1,
    "answer_train_max_len": 12,
    "eval_lengths": "4,8,12,24",
    "train_template_mode": "standard",
    "answer_train_template_mode": "standard",
    "eval_template_modes": "standard,paraphrase",
    "train_size": 4096,
    "answer_train_size": 2048,
    "eval_size": 256,
    "feature_batch_size": 8,
    "train_batch_size": 128,
    "eval_batch_size": 128,
    "bootstrap_steps": 1600,
    "answer_steps": 800,
    "eval_every": 800,
    "stage_eval_every": 800,
    "lr": 0.001,
    "answer_lr_scale": 0.1,
    "weight_decay": 0.01,
    "grad_clip": 1.0,
    "head_width": 768,
    "rank_temperature": 1.0,
    "arg_reader_mode": "after_op",
    "arg_window": 8,
    "arg_distance_temperature": 2.0,
    "trace_loss_weight": 1.0,
    "init_trace_loss_weight": 4.0,
    "op_trace_loss_weight": 1.0,
    "arg_trace_loss_weight": 4.0,
    "selection_loss_weight": 1.0,
    "role_pos_weight": 20.0,
    "role_count_loss_weight": 0.0,
    "init_selection_loss_weight": 1.0,
    "op_selection_loss_weight": 1.0,
    "arg_selection_loss_weight": 4.0,
    "executor_loss_weight": 1.0,
    "variants": "compiler_trace_then_answer,compiler_trace_then_answer_low_lr",
    "max_length": 512,
    "feature_seq_len": 320,
    "seed": 7,
    "output_dir": "experiments/qwen_shared_parser_compiler/runs/main_qwen35_after_op_retention",
    "checkpoint_dir": "",
    "verbose": false
  },
  "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": "AutoModelForMultimodalLM",
    "peft_installed": true,
    "gpu_name": "NVIDIA RTX 6000 Ada Generation",
    "gpu_vram_gb": 47.383
  },
  "hidden_dim": 2560,
  "dataset": {
    "bootstrap_train_size": 4096,
    "answer_train_size": 2048,
    "eval_size": 256,
    "bootstrap_train_lengths": [
      1,
      12
    ],
    "answer_train_lengths": [
      1,
      12
    ],
    "eval_lengths": [
      4,
      8,
      12,
      24
    ],
    "train_template_mode": "standard",
    "answer_train_template_mode": "standard",
    "eval_template_modes": [
      "standard",
      "paraphrase"
    ],
    "feature_seq_len": 320
  },
  "variants": {
    "compiler_trace_then_answer": {
      "variant": "compiler_trace_then_answer",
      "train_seconds": 238.11269402503967,
      "checkpoints": [
        "large_artifacts/qwen_shared_parser_compiler/checkpoints/main_qwen35_after_op_retention/compiler_trace_then_answer/checkpoint_final.pt"
      ],
      "train_log": [
        {
          "stage": "trace_bootstrap",
          "local_step": 1,
          "step": 1,
          "loss": 74.0871810913086,
          "trace_loss_active": true,
          "lr_scale": 1.0,
          "executor_loss": 4.579772472381592,
          "init_loss": 4.572800636291504,
          "op_loss": 1.0979887247085571,
          "arg_loss": 4.581509590148926,
          "init_attn_loss": 4.360878944396973,
          "op_attn_loss": 4.954745769500732,
          "arg_attn_loss": 4.843667507171631,
          "role_loss": 3.1018869876861572,
          "role_count_loss": 2.085042715072632,
          "standard_len4_direct_accuracy": NaN,
          "standard_len4_direct_target_mass": NaN,
          "standard_len4_executor_accuracy": 0.01953125,
          "standard_len4_executor_target_mass": 0.010314440354704857,
          "standard_len4_init_accuracy": 0.01171875,
          "standard_len4_init_pos_accuracy": 1.0,
          "standard_len4_op_accuracy": 0.333984375,
          "standard_len4_arg_accuracy": 0.0,
          "standard_len4_op_pos_accuracy": 0.0,
          "standard_len4_arg_pos_accuracy": 0.0,
          "standard_len4_program_exact": 0.0,
          "standard_len8_direct_accuracy": NaN,
          "standard_len8_direct_target_mass": NaN,
          "standard_len8_executor_accuracy": 0.02734375,
          "standard_len8_executor_target_mass": 0.010314198210835457,
          "standard_len8_init_accuracy": 0.01171875,
          "standard_len8_init_pos_accuracy": 1.0,
          "standard_len8_op_accuracy": 0.33251953125,
          "standard_len8_arg_accuracy": 0.02197265625,
          "standard_len8_op_pos_accuracy": 0.0,
          "standard_len8_arg_pos_accuracy": 0.0,
          "standard_len8_program_exact": 0.0,
          "standard_len12_direct_accuracy": NaN,
          "standard_len12_direct_target_mass": NaN,
          "standard_len12_executor_accuracy": 0.00390625,
          "standard_len12_executor_target_mass": 0.010299970395863056,
          "standard_len12_init_accuracy": 0.01171875,
          "standard_len12_init_pos_accuracy": 0.9921875,
          "standard_len12_op_accuracy": 0.3310546875,
          "standard_len12_arg_accuracy": 0.015950520833333332,
          "standard_len12_op_pos_accuracy": 0.0,
          "standard_len12_arg_pos_accuracy": 0.0,
          "standard_len12_program_exact": 0.0,
          "standard_len24_direct_accuracy": NaN,
          "standard_len24_direct_target_mass": NaN,
          "standard_len24_executor_accuracy": 0.015625,
          "standard_len24_executor_target_mass": 0.010273434221744537,
          "standard_len24_init_accuracy": 0.015625,
          "standard_len24_init_pos_accuracy": 0.99609375,
          "standard_len24_op_accuracy": 0.3367513020833333,
          "standard_len24_arg_accuracy": 0.019368489583333332,
          "standard_len24_op_pos_accuracy": 0.0,
          "standard_len24_arg_pos_accuracy": 0.0,
          "standard_len24_program_exact": 0.0,
          "paraphrase_len4_direct_accuracy": NaN,
          "paraphrase_len4_direct_target_mass": NaN,
          "paraphrase_len4_executor_accuracy": 0.0078125,
          "paraphrase_len4_executor_target_mass": 0.010329502634704113,
          "paraphrase_len4_init_accuracy": 0.01953125,
          "paraphrase_len4_init_pos_accuracy": 0.91796875,
          "paraphrase_len4_op_accuracy": 0.3642578125,
          "paraphrase_len4_arg_accuracy": 0.0,
          "paraphrase_len4_op_pos_accuracy": 0.0,
          "paraphrase_len4_arg_pos_accuracy": 0.0,
          "paraphrase_len4_program_exact": 0.0,
          "paraphrase_len8_direct_accuracy": NaN,
          "paraphrase_len8_direct_target_mass": NaN,
          "paraphrase_len8_executor_accuracy": 0.0078125,
          "paraphrase_len8_executor_target_mass": 0.010327577125281096,
          "paraphrase_len8_init_accuracy": 0.0078125,
          "paraphrase_len8_init_pos_accuracy": 0.90625,
          "paraphrase_len8_op_accuracy": 0.32470703125,
          "paraphrase_len8_arg_accuracy": 0.01123046875,
          "paraphrase_len8_op_pos_accuracy": 0.0,
          "paraphrase_len8_arg_pos_accuracy": 0.0,
          "paraphrase_len8_program_exact": 0.0,
          "paraphrase_len12_direct_accuracy": NaN,
          "paraphrase_len12_direct_target_mass": NaN,
          "paraphrase_len12_executor_accuracy": 0.0078125,
          "paraphrase_len12_executor_target_mass": 0.0102861188352108,
          "paraphrase_len12_init_accuracy": 0.0078125,
          "paraphrase_len12_init_pos_accuracy": 0.8984375,
          "paraphrase_len12_op_accuracy": 0.3216145833333333,
          "paraphrase_len12_arg_accuracy": 0.014322916666666666,
          "paraphrase_len12_op_pos_accuracy": 0.0,
          "paraphrase_len12_arg_pos_accuracy": 0.0,
          "paraphrase_len12_program_exact": 0.0,
          "paraphrase_len24_direct_accuracy": NaN,
          "paraphrase_len24_direct_target_mass": NaN,
          "paraphrase_len24_executor_accuracy": 0.00390625,
          "paraphrase_len24_executor_target_mass": 0.010341007728129625,
          "paraphrase_len24_init_accuracy": 0.01171875,
          "paraphrase_len24_init_pos_accuracy": 0.90234375,
          "paraphrase_len24_op_accuracy": 0.32763671875,
          "paraphrase_len24_arg_accuracy": 0.009602864583333334,
          "paraphrase_len24_op_pos_accuracy": 0.0,
          "paraphrase_len24_arg_pos_accuracy": 0.0,
          "paraphrase_len24_program_exact": 0.0
        },
        {
          "stage": "trace_bootstrap",
          "local_step": 800,
          "step": 800,
          "loss": 8.842819213867188,
          "trace_loss_active": true,
          "lr_scale": 1.0,
          "executor_loss": 3.315142869949341,
          "init_loss": 0.00013009343820158392,
          "op_loss": 0.31506526470184326,
          "arg_loss": 0.5490565896034241,
          "init_attn_loss": 0.00011863333929795772,
          "op_attn_loss": 0.5452795624732971,
          "arg_attn_loss": 0.5869467854499817,
          "role_loss": 0.12267852574586868,
          "role_count_loss": 0.1813075989484787,
          "standard_len4_direct_accuracy": NaN,
          "standard_len4_direct_target_mass": NaN,
          "standard_len4_executor_accuracy": 0.7265625,
          "standard_len4_executor_target_mass": 0.07960861921310425,
          "standard_len4_init_accuracy": 1.0,
          "standard_len4_init_pos_accuracy": 1.0,
          "standard_len4_op_accuracy": 1.0,
          "standard_len4_arg_accuracy": 0.927734375,
          "standard_len4_op_pos_accuracy": 1.0,
          "standard_len4_arg_pos_accuracy": 0.9560546875,
          "standard_len4_program_exact": 0.7265625,
          "standard_len8_direct_accuracy": NaN,
          "standard_len8_direct_target_mass": NaN,
          "standard_len8_executor_accuracy": 0.5703125,
          "standard_len8_executor_target_mass": 0.012177260592579842,
          "standard_len8_init_accuracy": 1.0,
          "standard_len8_init_pos_accuracy": 1.0,
          "standard_len8_op_accuracy": 0.9970703125,
          "standard_len8_arg_accuracy": 0.93408203125,
          "standard_len8_op_pos_accuracy": 1.0,
          "standard_len8_arg_pos_accuracy": 0.94677734375,
          "standard_len8_program_exact": 0.56640625,
          "standard_len12_direct_accuracy": NaN,
          "standard_len12_direct_target_mass": NaN,
          "standard_len12_executor_accuracy": 0.328125,
          "standard_len12_executor_target_mass": 0.010367827024310827,
          "standard_len12_init_accuracy": 1.0,
          "standard_len12_init_pos_accuracy": 1.0,
          "standard_len12_op_accuracy": 0.99609375,
          "standard_len12_arg_accuracy": 0.9189453125,
          "standard_len12_op_pos_accuracy": 1.0,
          "standard_len12_arg_pos_accuracy": 0.9375,
          "standard_len12_program_exact": 0.32421875,
          "standard_len24_direct_accuracy": NaN,
          "standard_len24_direct_target_mass": NaN,
          "standard_len24_executor_accuracy": 0.04296875,
          "standard_len24_executor_target_mass": 0.010309273842722178,
          "standard_len24_init_accuracy": 1.0,
          "standard_len24_init_pos_accuracy": 1.0,
          "standard_len24_op_accuracy": 0.9762369791666666,
          "standard_len24_arg_accuracy": 0.908203125,
          "standard_len24_op_pos_accuracy": 1.0,
          "standard_len24_arg_pos_accuracy": 0.9366861979166666,
          "standard_len24_program_exact": 0.03515625,
          "paraphrase_len4_direct_accuracy": NaN,
          "paraphrase_len4_direct_target_mass": NaN,
          "paraphrase_len4_executor_accuracy": 0.05859375,
          "paraphrase_len4_executor_target_mass": 0.014704071916639805,
          "paraphrase_len4_init_accuracy": 0.73828125,
          "paraphrase_len4_init_pos_accuracy": 0.91796875,
          "paraphrase_len4_op_accuracy": 0.6669921875,
          "paraphrase_len4_arg_accuracy": 0.5048828125,
          "paraphrase_len4_op_pos_accuracy": 0.576171875,
          "paraphrase_len4_arg_pos_accuracy": 0.560546875,
          "paraphrase_len4_program_exact": 0.0234375,
          "paraphrase_len8_direct_accuracy": NaN,
          "paraphrase_len8_direct_target_mass": NaN,
          "paraphrase_len8_executor_accuracy": 0.01171875,
          "paraphrase_len8_executor_target_mass": 0.010036191437393427,
          "paraphrase_len8_init_accuracy": 0.74609375,
          "paraphrase_len8_init_pos_accuracy": 0.921875,
          "paraphrase_len8_op_accuracy": 0.5654296875,
          "paraphrase_len8_arg_accuracy": 0.36767578125,
          "paraphrase_len8_op_pos_accuracy": 0.375,
          "paraphrase_len8_arg_pos_accuracy": 0.36328125,
          "paraphrase_len8_program_exact": 0.0,
          "paraphrase_len12_direct_accuracy": NaN,
          "paraphrase_len12_direct_target_mass": NaN,
          "paraphrase_len12_executor_accuracy": 0.01171875,
          "paraphrase_len12_executor_target_mass": 0.010314361657947302,
          "paraphrase_len12_init_accuracy": 0.7421875,
          "paraphrase_len12_init_pos_accuracy": 0.921875,
          "paraphrase_len12_op_accuracy": 0.5162760416666666,
          "paraphrase_len12_arg_accuracy": 0.2916666666666667,
          "paraphrase_len12_op_pos_accuracy": 0.2672526041666667,
          "paraphrase_len12_arg_pos_accuracy": 0.2708333333333333,
          "paraphrase_len12_program_exact": 0.0,
          "paraphrase_len24_direct_accuracy": NaN,
          "paraphrase_len24_direct_target_mass": NaN,
          "paraphrase_len24_executor_accuracy": 0.01953125,
          "paraphrase_len24_executor_target_mass": 0.01030915230512619,
          "paraphrase_len24_init_accuracy": 0.71484375,
          "paraphrase_len24_init_pos_accuracy": 0.921875,
          "paraphrase_len24_op_accuracy": 0.4361979166666667,
          "paraphrase_len24_arg_accuracy": 0.1640625,
          "paraphrase_len24_op_pos_accuracy": 0.12825520833333334,
          "paraphrase_len24_arg_pos_accuracy": 0.13297526041666666,
          "paraphrase_len24_program_exact": 0.0
        },
        {
          "stage": "trace_bootstrap",
          "local_step": 1600,
          "step": 1600,
          "loss": 8.43596076965332,
          "trace_loss_active": true,
          "lr_scale": 1.0,
          "executor_loss": 3.0295944213867188,
          "init_loss": 6.301762186922133e-05,
          "op_loss": 0.2963768541812897,
          "arg_loss": 0.5254144072532654,
          "init_attn_loss": 0.00018633034778758883,
          "op_attn_loss": 0.5555647015571594,
          "arg_attn_loss": 0.5689468383789062,
          "role_loss": 0.17654094099998474,
          "role_count_loss": 0.20004932582378387,
          "standard_len4_direct_accuracy": NaN,
          "standard_len4_direct_target_mass": NaN,
          "standard_len4_executor_accuracy": 0.73828125,
          "standard_len4_executor_target_mass": 0.08165208995342255,
          "standard_len4_init_accuracy": 1.0,
          "standard_len4_init_pos_accuracy": 1.0,
          "standard_len4_op_accuracy": 0.994140625,
          "standard_len4_arg_accuracy": 0.9326171875,
          "standard_len4_op_pos_accuracy": 1.0,
          "standard_len4_arg_pos_accuracy": 0.94140625,
          "standard_len4_program_exact": 0.73828125,
          "standard_len8_direct_accuracy": NaN,
          "standard_len8_direct_target_mass": NaN,
          "standard_len8_executor_accuracy": 0.51171875,
          "standard_len8_executor_target_mass": 0.01230234932154417,
          "standard_len8_init_accuracy": 1.0,
          "standard_len8_init_pos_accuracy": 1.0,
          "standard_len8_op_accuracy": 0.974609375,
          "standard_len8_arg_accuracy": 0.93798828125,
          "standard_len8_op_pos_accuracy": 1.0,
          "standard_len8_arg_pos_accuracy": 0.94580078125,
          "standard_len8_program_exact": 0.5078125,
          "standard_len12_direct_accuracy": NaN,
          "standard_len12_direct_target_mass": NaN,
          "standard_len12_executor_accuracy": 0.2421875,
          "standard_len12_executor_target_mass": 0.010357526130974293,
          "standard_len12_init_accuracy": 1.0,
          "standard_len12_init_pos_accuracy": 1.0,
          "standard_len12_op_accuracy": 0.9680989583333334,
          "standard_len12_arg_accuracy": 0.9114583333333334,
          "standard_len12_op_pos_accuracy": 0.9990234375,
          "standard_len12_arg_pos_accuracy": 0.9283854166666666,
          "standard_len12_program_exact": 0.23828125,
          "standard_len24_direct_accuracy": NaN,
          "standard_len24_direct_target_mass": NaN,
          "standard_len24_executor_accuracy": 0.01171875,
          "standard_len24_executor_target_mass": 0.010309250093996525,
          "standard_len24_init_accuracy": 1.0,
          "standard_len24_init_pos_accuracy": 1.0,
          "standard_len24_op_accuracy": 0.7403971354166666,
          "standard_len24_arg_accuracy": 0.6017252604166666,
          "standard_len24_op_pos_accuracy": 0.61376953125,
          "standard_len24_arg_pos_accuracy": 0.5831705729166666,
          "standard_len24_program_exact": 0.0,
          "paraphrase_len4_direct_accuracy": NaN,
          "paraphrase_len4_direct_target_mass": NaN,
          "paraphrase_len4_executor_accuracy": 0.0234375,
          "paraphrase_len4_executor_target_mass": 0.01477328222244978,
          "paraphrase_len4_init_accuracy": 0.55078125,
          "paraphrase_len4_init_pos_accuracy": 0.8359375,
          "paraphrase_len4_op_accuracy": 0.6640625,
          "paraphrase_len4_arg_accuracy": 0.541015625,
          "paraphrase_len4_op_pos_accuracy": 0.5927734375,
          "paraphrase_len4_arg_pos_accuracy": 0.57421875,
          "paraphrase_len4_program_exact": 0.015625,
          "paraphrase_len8_direct_accuracy": NaN,
          "paraphrase_len8_direct_target_mass": NaN,
          "paraphrase_len8_executor_accuracy": 0.00390625,
          "paraphrase_len8_executor_target_mass": 0.010181421414017677,
          "paraphrase_len8_init_accuracy": 0.53515625,
          "paraphrase_len8_init_pos_accuracy": 0.8203125,
          "paraphrase_len8_op_accuracy": 0.580078125,
          "paraphrase_len8_arg_accuracy": 0.3896484375,
          "paraphrase_len8_op_pos_accuracy": 0.41162109375,
          "paraphrase_len8_arg_pos_accuracy": 0.39599609375,
          "paraphrase_len8_program_exact": 0.0,
          "paraphrase_len12_direct_accuracy": NaN,
          "paraphrase_len12_direct_target_mass": NaN,
          "paraphrase_len12_executor_accuracy": 0.0078125,
          "paraphrase_len12_executor_target_mass": 0.010312501806765795,
          "paraphrase_len12_init_accuracy": 0.53125,
          "paraphrase_len12_init_pos_accuracy": 0.8515625,
          "paraphrase_len12_op_accuracy": 0.5211588541666666,
          "paraphrase_len12_arg_accuracy": 0.3069661458333333,
          "paraphrase_len12_op_pos_accuracy": 0.3043619791666667,
          "paraphrase_len12_arg_pos_accuracy": 0.3046875,
          "paraphrase_len12_program_exact": 0.0,
          "paraphrase_len24_direct_accuracy": NaN,
          "paraphrase_len24_direct_target_mass": NaN,
          "paraphrase_len24_executor_accuracy": 0.00390625,
          "paraphrase_len24_executor_target_mass": 0.010308923199772835,
          "paraphrase_len24_init_accuracy": 0.4921875,
          "paraphrase_len24_init_pos_accuracy": 0.85546875,
          "paraphrase_len24_op_accuracy": 0.4327799479166667,
          "paraphrase_len24_arg_accuracy": 0.17936197916666666,
          "paraphrase_len24_op_pos_accuracy": 0.15690104166666666,
          "paraphrase_len24_arg_pos_accuracy": 0.15576171875,
          "paraphrase_len24_program_exact": 0.0
        },
        {
          "stage": "answer_retention",
          "local_step": 1,
          "step": 1601,
          "loss": 3.3286516666412354,
          "trace_loss_active": false,
          "lr_scale": 1.0,
          "executor_loss": 3.3286516666412354,
          "standard_len4_direct_accuracy": NaN,
          "standard_len4_direct_target_mass": NaN,
          "standard_len4_executor_accuracy": 0.69140625,
          "standard_len4_executor_target_mass": 0.0798642747104168,
          "standard_len4_init_accuracy": 1.0,
          "standard_len4_init_pos_accuracy": 1.0,
          "standard_len4_op_accuracy": 0.9814453125,
          "standard_len4_arg_accuracy": 0.9287109375,
          "standard_len4_op_pos_accuracy": 1.0,
          "standard_len4_arg_pos_accuracy": 0.94921875,
          "standard_len4_program_exact": 0.69140625,
          "standard_len8_direct_accuracy": NaN,
          "standard_len8_direct_target_mass": NaN,
          "standard_len8_executor_accuracy": 0.41015625,
          "standard_len8_executor_target_mass": 0.012099228333681822,
          "standard_len8_init_accuracy": 1.0,
          "standard_len8_init_pos_accuracy": 1.0,
          "standard_len8_op_accuracy": 0.962890625,
          "standard_len8_arg_accuracy": 0.92529296875,
          "standard_len8_op_pos_accuracy": 1.0,
          "standard_len8_arg_pos_accuracy": 0.94775390625,
          "standard_len8_program_exact": 0.40625,
          "standard_len12_direct_accuracy": NaN,
          "standard_len12_direct_target_mass": NaN,
          "standard_len12_executor_accuracy": 0.15234375,
          "standard_len12_executor_target_mass": 0.010355750564485788,
          "standard_len12_init_accuracy": 1.0,
          "standard_len12_init_pos_accuracy": 1.0,
          "standard_len12_op_accuracy": 0.9475911458333334,
          "standard_len12_arg_accuracy": 0.890625,
          "standard_len12_op_pos_accuracy": 0.9957682291666666,
          "standard_len12_arg_pos_accuracy": 0.9241536458333334,
          "standard_len12_program_exact": 0.14453125,
          "standard_len24_direct_accuracy": NaN,
          "standard_len24_direct_target_mass": NaN,
          "standard_len24_executor_accuracy": 0.0078125,
          "standard_len24_executor_target_mass": 0.010309269186109304,
          "standard_len24_init_accuracy": 1.0,
          "standard_len24_init_pos_accuracy": 1.0,
          "standard_len24_op_accuracy": 0.7117513020833334,
          "standard_len24_arg_accuracy": 0.5623372395833334,
          "standard_len24_op_pos_accuracy": 0.5768229166666666,
          "standard_len24_arg_pos_accuracy": 0.54443359375,
          "standard_len24_program_exact": 0.0,
          "paraphrase_len4_direct_accuracy": NaN,
          "paraphrase_len4_direct_target_mass": NaN,
          "paraphrase_len4_executor_accuracy": 0.03125,
          "paraphrase_len4_executor_target_mass": 0.014792872127145529,
          "paraphrase_len4_init_accuracy": 0.52734375,
          "paraphrase_len4_init_pos_accuracy": 0.8359375,
          "paraphrase_len4_op_accuracy": 0.6708984375,
          "paraphrase_len4_arg_accuracy": 0.5439453125,
          "paraphrase_len4_op_pos_accuracy": 0.6025390625,
          "paraphrase_len4_arg_pos_accuracy": 0.5830078125,
          "paraphrase_len4_program_exact": 0.015625,
          "paraphrase_len8_direct_accuracy": NaN,
          "paraphrase_len8_direct_target_mass": NaN,
          "paraphrase_len8_executor_accuracy": 0.0,
          "paraphrase_len8_executor_target_mass": 0.010128918569535017,
          "paraphrase_len8_init_accuracy": 0.5,
          "paraphrase_len8_init_pos_accuracy": 0.80859375,
          "paraphrase_len8_op_accuracy": 0.5888671875,
          "paraphrase_len8_arg_accuracy": 0.39697265625,
          "paraphrase_len8_op_pos_accuracy": 0.43359375,
          "paraphrase_len8_arg_pos_accuracy": 0.4130859375,
          "paraphrase_len8_program_exact": 0.0,
          "paraphrase_len12_direct_accuracy": NaN,
          "paraphrase_len12_direct_target_mass": NaN,
          "paraphrase_len12_executor_accuracy": 0.01171875,
          "paraphrase_len12_executor_target_mass": 0.010327024850994349,
          "paraphrase_len12_init_accuracy": 0.5234375,
          "paraphrase_len12_init_pos_accuracy": 0.8515625,
          "paraphrase_len12_op_accuracy": 0.5325520833333334,
          "paraphrase_len12_arg_accuracy": 0.3154296875,
          "paraphrase_len12_op_pos_accuracy": 0.326171875,
          "paraphrase_len12_arg_pos_accuracy": 0.3151041666666667,
          "paraphrase_len12_program_exact": 0.0,
          "paraphrase_len24_direct_accuracy": NaN,
          "paraphrase_len24_direct_target_mass": NaN,
          "paraphrase_len24_executor_accuracy": 0.00390625,
          "paraphrase_len24_executor_target_mass": 0.010308966040611267,
          "paraphrase_len24_init_accuracy": 0.48046875,
          "paraphrase_len24_init_pos_accuracy": 0.84375,
          "paraphrase_len24_op_accuracy": 0.4461263020833333,
          "paraphrase_len24_arg_accuracy": 0.1962890625,
          "paraphrase_len24_op_pos_accuracy": 0.17203776041666666,
          "paraphrase_len24_arg_pos_accuracy": 0.17073567708333334,
          "paraphrase_len24_program_exact": 0.0
        },
        {
          "stage": "answer_retention",
          "local_step": 800,
          "step": 2400,
          "loss": 2.9209001064300537,
          "trace_loss_active": false,
          "lr_scale": 1.0,
          "executor_loss": 2.9209001064300537,
          "standard_len4_direct_accuracy": NaN,
          "standard_len4_direct_target_mass": NaN,
          "standard_len4_executor_accuracy": 0.6015625,
          "standard_len4_executor_target_mass": 0.1518644094467163,
          "standard_len4_init_accuracy": 1.0,
          "standard_len4_init_pos_accuracy": 0.9921875,
          "standard_len4_op_accuracy": 0.9521484375,
          "standard_len4_arg_accuracy": 0.908203125,
          "standard_len4_op_pos_accuracy": 0.361328125,
          "standard_len4_arg_pos_accuracy": 0.4755859375,
          "standard_len4_program_exact": 0.6015625,
          "standard_len8_direct_accuracy": NaN,
          "standard_len8_direct_target_mass": NaN,
          "standard_len8_executor_accuracy": 0.00390625,
          "standard_len8_executor_target_mass": 0.010400094091892242,
          "standard_len8_init_accuracy": 1.0,
          "standard_len8_init_pos_accuracy": 0.9921875,
          "standard_len8_op_accuracy": 0.78759765625,
          "standard_len8_arg_accuracy": 0.69482421875,
          "standard_len8_op_pos_accuracy": 0.18505859375,
          "standard_len8_arg_pos_accuracy": 0.2431640625,
          "standard_len8_program_exact": 0.0,
          "standard_len12_direct_accuracy": NaN,
          "standard_len12_direct_target_mass": NaN,
          "standard_len12_executor_accuracy": 0.01171875,
          "standard_len12_executor_target_mass": 0.010485976003110409,
          "standard_len12_init_accuracy": 1.0,
          "standard_len12_init_pos_accuracy": 0.9921875,
          "standard_len12_op_accuracy": 0.6806640625,
          "standard_len12_arg_accuracy": 0.5504557291666666,
          "standard_len12_op_pos_accuracy": 0.12141927083333333,
          "standard_len12_arg_pos_accuracy": 0.16373697916666666,
          "standard_len12_program_exact": 0.0,
          "standard_len24_direct_accuracy": NaN,
          "standard_len24_direct_target_mass": NaN,
          "standard_len24_executor_accuracy": 0.0078125,
          "standard_len24_executor_target_mass": 0.010305849835276604,
          "standard_len24_init_accuracy": 1.0,
          "standard_len24_init_pos_accuracy": 0.96484375,
          "standard_len24_op_accuracy": 0.5540364583333334,
          "standard_len24_arg_accuracy": 0.36279296875,
          "standard_len24_op_pos_accuracy": 0.06363932291666667,
          "standard_len24_arg_pos_accuracy": 0.08089192708333333,
          "standard_len24_program_exact": 0.0,
          "paraphrase_len4_direct_accuracy": NaN,
          "paraphrase_len4_direct_target_mass": NaN,
          "paraphrase_len4_executor_accuracy": 0.03125,
          "paraphrase_len4_executor_target_mass": 0.014552258886396885,
          "paraphrase_len4_init_accuracy": 0.4921875,
          "paraphrase_len4_init_pos_accuracy": 0.62890625,
          "paraphrase_len4_op_accuracy": 0.6171875,
          "paraphrase_len4_arg_accuracy": 0.443359375,
          "paraphrase_len4_op_pos_accuracy": 0.08984375,
          "paraphrase_len4_arg_pos_accuracy": 0.3359375,
          "paraphrase_len4_program_exact": 0.0078125,
          "paraphrase_len8_direct_accuracy": NaN,
          "paraphrase_len8_direct_target_mass": NaN,
          "paraphrase_len8_executor_accuracy": 0.01953125,
          "paraphrase_len8_executor_target_mass": 0.010807174257934093,
          "paraphrase_len8_init_accuracy": 0.44921875,
          "paraphrase_len8_init_pos_accuracy": 0.578125,
          "paraphrase_len8_op_accuracy": 0.50927734375,
          "paraphrase_len8_arg_accuracy": 0.30712890625,
          "paraphrase_len8_op_pos_accuracy": 0.06884765625,
          "paraphrase_len8_arg_pos_accuracy": 0.203125,
          "paraphrase_len8_program_exact": 0.0,
          "paraphrase_len12_direct_accuracy": NaN,
          "paraphrase_len12_direct_target_mass": NaN,
          "paraphrase_len12_executor_accuracy": 0.01953125,
          "paraphrase_len12_executor_target_mass": 0.010312123224139214,
          "paraphrase_len12_init_accuracy": 0.4453125,
          "paraphrase_len12_init_pos_accuracy": 0.640625,
          "paraphrase_len12_op_accuracy": 0.470703125,
          "paraphrase_len12_arg_accuracy": 0.24739583333333334,
          "paraphrase_len12_op_pos_accuracy": 0.050455729166666664,
          "paraphrase_len12_arg_pos_accuracy": 0.14420572916666666,
          "paraphrase_len12_program_exact": 0.0,
          "paraphrase_len24_direct_accuracy": NaN,
          "paraphrase_len24_direct_target_mass": NaN,
          "paraphrase_len24_executor_accuracy": 0.01171875,
          "paraphrase_len24_executor_target_mass": 0.01031398680061102,
          "paraphrase_len24_init_accuracy": 0.390625,
          "paraphrase_len24_init_pos_accuracy": 0.54296875,
          "paraphrase_len24_op_accuracy": 0.4371744791666667,
          "paraphrase_len24_arg_accuracy": 0.21842447916666666,
          "paraphrase_len24_op_pos_accuracy": 0.0859375,
          "paraphrase_len24_arg_pos_accuracy": 0.13525390625,
          "paraphrase_len24_program_exact": 0.0
        }
      ],
      "final_metrics": {
        "standard_len4": {
          "variant": "compiler_trace_then_answer",
          "n": 256.0,
          "direct_accuracy": NaN,
          "direct_target_mass": NaN,
          "executor_accuracy": 0.6015625,
          "executor_target_mass": 0.1518644094467163,
          "init_accuracy": 1.0,
          "init_pos_accuracy": 0.9921875,
          "op_accuracy": 0.9521484375,
          "arg_accuracy": 0.908203125,
          "op_pos_accuracy": 0.361328125,
          "arg_pos_accuracy": 0.4755859375,
          "program_exact": 0.6015625
        },
        "standard_len8": {
          "variant": "compiler_trace_then_answer",
          "n": 256.0,
          "direct_accuracy": NaN,
          "direct_target_mass": NaN,
          "executor_accuracy": 0.00390625,
          "executor_target_mass": 0.010400094091892242,
          "init_accuracy": 1.0,
          "init_pos_accuracy": 0.9921875,
          "op_accuracy": 0.78759765625,
          "arg_accuracy": 0.69482421875,
          "op_pos_accuracy": 0.18505859375,
          "arg_pos_accuracy": 0.2431640625,
          "program_exact": 0.0
        },
        "standard_len12": {
          "variant": "compiler_trace_then_answer",
          "n": 256.0,
          "direct_accuracy": NaN,
          "direct_target_mass": NaN,
          "executor_accuracy": 0.01171875,
          "executor_target_mass": 0.010485976003110409,
          "init_accuracy": 1.0,
          "init_pos_accuracy": 0.9921875,
          "op_accuracy": 0.6806640625,
          "arg_accuracy": 0.5504557291666666,
          "op_pos_accuracy": 0.12141927083333333,
          "arg_pos_accuracy": 0.16373697916666666,
          "program_exact": 0.0
        },
        "standard_len24": {
          "variant": "compiler_trace_then_answer",
          "n": 256.0,
          "direct_accuracy": NaN,
          "direct_target_mass": NaN,
          "executor_accuracy": 0.0078125,
          "executor_target_mass": 0.010305849835276604,
          "init_accuracy": 1.0,
          "init_pos_accuracy": 0.96484375,
          "op_accuracy": 0.5540364583333334,
          "arg_accuracy": 0.36279296875,
          "op_pos_accuracy": 0.06363932291666667,
          "arg_pos_accuracy": 0.08089192708333333,
          "program_exact": 0.0
        },
        "paraphrase_len4": {
          "variant": "compiler_trace_then_answer",
          "n": 256.0,
          "direct_accuracy": NaN,
          "direct_target_mass": NaN,
          "executor_accuracy": 0.03125,
          "executor_target_mass": 0.014552258886396885,
          "init_accuracy": 0.4921875,
          "init_pos_accuracy": 0.62890625,
          "op_accuracy": 0.6171875,
          "arg_accuracy": 0.443359375,
          "op_pos_accuracy": 0.08984375,
          "arg_pos_accuracy": 0.3359375,
          "program_exact": 0.0078125
        },
        "paraphrase_len8": {
          "variant": "compiler_trace_then_answer",
          "n": 256.0,
          "direct_accuracy": NaN,
          "direct_target_mass": NaN,
          "executor_accuracy": 0.01953125,
          "executor_target_mass": 0.010807174257934093,
          "init_accuracy": 0.44921875,
          "init_pos_accuracy": 0.578125,
          "op_accuracy": 0.50927734375,
          "arg_accuracy": 0.30712890625,
          "op_pos_accuracy": 0.06884765625,
          "arg_pos_accuracy": 0.203125,
          "program_exact": 0.0
        },
        "paraphrase_len12": {
          "variant": "compiler_trace_then_answer",
          "n": 256.0,
          "direct_accuracy": NaN,
          "direct_target_mass": NaN,
          "executor_accuracy": 0.01953125,
          "executor_target_mass": 0.010312123224139214,
          "init_accuracy": 0.4453125,
          "init_pos_accuracy": 0.640625,
          "op_accuracy": 0.470703125,
          "arg_accuracy": 0.24739583333333334,
          "op_pos_accuracy": 0.050455729166666664,
          "arg_pos_accuracy": 0.14420572916666666,
          "program_exact": 0.0
        },
        "paraphrase_len24": {
          "variant": "compiler_trace_then_answer",
          "n": 256.0,
          "direct_accuracy": NaN,
          "direct_target_mass": NaN,
          "executor_accuracy": 0.01171875,
          "executor_target_mass": 0.01031398680061102,
          "init_accuracy": 0.390625,
          "init_pos_accuracy": 0.54296875,
          "op_accuracy": 0.4371744791666667,
          "arg_accuracy": 0.21842447916666666,
          "op_pos_accuracy": 0.0859375,
          "arg_pos_accuracy": 0.13525390625,
          "program_exact": 0.0
        }
      }
    },
    "compiler_trace_then_answer_low_lr": {
      "variant": "compiler_trace_then_answer_low_lr",
      "train_seconds": 238.32915544509888,
      "checkpoints": [
        "large_artifacts/qwen_shared_parser_compiler/checkpoints/main_qwen35_after_op_retention/compiler_trace_then_answer_low_lr/checkpoint_final.pt"
      ],
      "train_log": [
        {
          "stage": "trace_bootstrap",
          "local_step": 1,
          "step": 1,
          "loss": 74.0871810913086,
          "trace_loss_active": true,
          "lr_scale": 1.0,
          "executor_loss": 4.579772472381592,
          "init_loss": 4.572800636291504,
          "op_loss": 1.0979887247085571,
          "arg_loss": 4.581509590148926,
          "init_attn_loss": 4.360878944396973,
          "op_attn_loss": 4.954745769500732,
          "arg_attn_loss": 4.843667507171631,
          "role_loss": 3.1018869876861572,
          "role_count_loss": 2.085042715072632,
          "standard_len4_direct_accuracy": NaN,
          "standard_len4_direct_target_mass": NaN,
          "standard_len4_executor_accuracy": 0.01953125,
          "standard_len4_executor_target_mass": 0.010314440354704857,
          "standard_len4_init_accuracy": 0.01171875,
          "standard_len4_init_pos_accuracy": 1.0,
          "standard_len4_op_accuracy": 0.333984375,
          "standard_len4_arg_accuracy": 0.0,
          "standard_len4_op_pos_accuracy": 0.0,
          "standard_len4_arg_pos_accuracy": 0.0,
          "standard_len4_program_exact": 0.0,
          "standard_len8_direct_accuracy": NaN,
          "standard_len8_direct_target_mass": NaN,
          "standard_len8_executor_accuracy": 0.02734375,
          "standard_len8_executor_target_mass": 0.010314198210835457,
          "standard_len8_init_accuracy": 0.01171875,
          "standard_len8_init_pos_accuracy": 1.0,
          "standard_len8_op_accuracy": 0.33251953125,
          "standard_len8_arg_accuracy": 0.02197265625,
          "standard_len8_op_pos_accuracy": 0.0,
          "standard_len8_arg_pos_accuracy": 0.0,
          "standard_len8_program_exact": 0.0,
          "standard_len12_direct_accuracy": NaN,
          "standard_len12_direct_target_mass": NaN,
          "standard_len12_executor_accuracy": 0.00390625,
          "standard_len12_executor_target_mass": 0.010299970395863056,
          "standard_len12_init_accuracy": 0.01171875,
          "standard_len12_init_pos_accuracy": 0.9921875,
          "standard_len12_op_accuracy": 0.3310546875,
          "standard_len12_arg_accuracy": 0.015950520833333332,
          "standard_len12_op_pos_accuracy": 0.0,
          "standard_len12_arg_pos_accuracy": 0.0,
          "standard_len12_program_exact": 0.0,
          "standard_len24_direct_accuracy": NaN,
          "standard_len24_direct_target_mass": NaN,
          "standard_len24_executor_accuracy": 0.015625,
          "standard_len24_executor_target_mass": 0.010273434221744537,
          "standard_len24_init_accuracy": 0.015625,
          "standard_len24_init_pos_accuracy": 0.99609375,
          "standard_len24_op_accuracy": 0.3367513020833333,
          "standard_len24_arg_accuracy": 0.019368489583333332,
          "standard_len24_op_pos_accuracy": 0.0,
          "standard_len24_arg_pos_accuracy": 0.0,
          "standard_len24_program_exact": 0.0,
          "paraphrase_len4_direct_accuracy": NaN,
          "paraphrase_len4_direct_target_mass": NaN,
          "paraphrase_len4_executor_accuracy": 0.0078125,
          "paraphrase_len4_executor_target_mass": 0.010329502634704113,
          "paraphrase_len4_init_accuracy": 0.01953125,
          "paraphrase_len4_init_pos_accuracy": 0.91796875,
          "paraphrase_len4_op_accuracy": 0.3642578125,
          "paraphrase_len4_arg_accuracy": 0.0,
          "paraphrase_len4_op_pos_accuracy": 0.0,
          "paraphrase_len4_arg_pos_accuracy": 0.0,
          "paraphrase_len4_program_exact": 0.0,
          "paraphrase_len8_direct_accuracy": NaN,
          "paraphrase_len8_direct_target_mass": NaN,
          "paraphrase_len8_executor_accuracy": 0.0078125,
          "paraphrase_len8_executor_target_mass": 0.010327577125281096,
          "paraphrase_len8_init_accuracy": 0.0078125,
          "paraphrase_len8_init_pos_accuracy": 0.90625,
          "paraphrase_len8_op_accuracy": 0.32470703125,
          "paraphrase_len8_arg_accuracy": 0.01123046875,
          "paraphrase_len8_op_pos_accuracy": 0.0,
          "paraphrase_len8_arg_pos_accuracy": 0.0,
          "paraphrase_len8_program_exact": 0.0,
          "paraphrase_len12_direct_accuracy": NaN,
          "paraphrase_len12_direct_target_mass": NaN,
          "paraphrase_len12_executor_accuracy": 0.0078125,
          "paraphrase_len12_executor_target_mass": 0.0102861188352108,
          "paraphrase_len12_init_accuracy": 0.0078125,
          "paraphrase_len12_init_pos_accuracy": 0.8984375,
          "paraphrase_len12_op_accuracy": 0.3216145833333333,
          "paraphrase_len12_arg_accuracy": 0.014322916666666666,
          "paraphrase_len12_op_pos_accuracy": 0.0,
          "paraphrase_len12_arg_pos_accuracy": 0.0,
          "paraphrase_len12_program_exact": 0.0,
          "paraphrase_len24_direct_accuracy": NaN,
          "paraphrase_len24_direct_target_mass": NaN,
          "paraphrase_len24_executor_accuracy": 0.00390625,
          "paraphrase_len24_executor_target_mass": 0.010341007728129625,
          "paraphrase_len24_init_accuracy": 0.01171875,
          "paraphrase_len24_init_pos_accuracy": 0.90234375,
          "paraphrase_len24_op_accuracy": 0.32763671875,
          "paraphrase_len24_arg_accuracy": 0.009602864583333334,
          "paraphrase_len24_op_pos_accuracy": 0.0,
          "paraphrase_len24_arg_pos_accuracy": 0.0,
          "paraphrase_len24_program_exact": 0.0
        },
        {
          "stage": "trace_bootstrap",
          "local_step": 800,
          "step": 800,
          "loss": 8.842819213867188,
          "trace_loss_active": true,
          "lr_scale": 1.0,
          "executor_loss": 3.315142869949341,
          "init_loss": 0.00013009343820158392,
          "op_loss": 0.31506526470184326,
          "arg_loss": 0.5490565896034241,
          "init_attn_loss": 0.00011863333929795772,
          "op_attn_loss": 0.5452795624732971,
          "arg_attn_loss": 0.5869467854499817,
          "role_loss": 0.12267852574586868,
          "role_count_loss": 0.1813075989484787,
          "standard_len4_direct_accuracy": NaN,
          "standard_len4_direct_target_mass": NaN,
          "standard_len4_executor_accuracy": 0.7265625,
          "standard_len4_executor_target_mass": 0.07960861921310425,
          "standard_len4_init_accuracy": 1.0,
          "standard_len4_init_pos_accuracy": 1.0,
          "standard_len4_op_accuracy": 1.0,
          "standard_len4_arg_accuracy": 0.927734375,
          "standard_len4_op_pos_accuracy": 1.0,
          "standard_len4_arg_pos_accuracy": 0.9560546875,
          "standard_len4_program_exact": 0.7265625,
          "standard_len8_direct_accuracy": NaN,
          "standard_len8_direct_target_mass": NaN,
          "standard_len8_executor_accuracy": 0.5703125,
          "standard_len8_executor_target_mass": 0.012177260592579842,
          "standard_len8_init_accuracy": 1.0,
          "standard_len8_init_pos_accuracy": 1.0,
          "standard_len8_op_accuracy": 0.9970703125,
          "standard_len8_arg_accuracy": 0.93408203125,
          "standard_len8_op_pos_accuracy": 1.0,
          "standard_len8_arg_pos_accuracy": 0.94677734375,
          "standard_len8_program_exact": 0.56640625,
          "standard_len12_direct_accuracy": NaN,
          "standard_len12_direct_target_mass": NaN,
          "standard_len12_executor_accuracy": 0.328125,
          "standard_len12_executor_target_mass": 0.010367827024310827,
          "standard_len12_init_accuracy": 1.0,
          "standard_len12_init_pos_accuracy": 1.0,
          "standard_len12_op_accuracy": 0.99609375,
          "standard_len12_arg_accuracy": 0.9189453125,
          "standard_len12_op_pos_accuracy": 1.0,
          "standard_len12_arg_pos_accuracy": 0.9375,
          "standard_len12_program_exact": 0.32421875,
          "standard_len24_direct_accuracy": NaN,
          "standard_len24_direct_target_mass": NaN,
          "standard_len24_executor_accuracy": 0.04296875,
          "standard_len24_executor_target_mass": 0.010309273842722178,
          "standard_len24_init_accuracy": 1.0,
          "standard_len24_init_pos_accuracy": 1.0,
          "standard_len24_op_accuracy": 0.9762369791666666,
          "standard_len24_arg_accuracy": 0.908203125,
          "standard_len24_op_pos_accuracy": 1.0,
          "standard_len24_arg_pos_accuracy": 0.9366861979166666,
          "standard_len24_program_exact": 0.03515625,
          "paraphrase_len4_direct_accuracy": NaN,
          "paraphrase_len4_direct_target_mass": NaN,
          "paraphrase_len4_executor_accuracy": 0.05859375,
          "paraphrase_len4_executor_target_mass": 0.014704071916639805,
          "paraphrase_len4_init_accuracy": 0.73828125,
          "paraphrase_len4_init_pos_accuracy": 0.91796875,
          "paraphrase_len4_op_accuracy": 0.6669921875,
          "paraphrase_len4_arg_accuracy": 0.5048828125,
          "paraphrase_len4_op_pos_accuracy": 0.576171875,
          "paraphrase_len4_arg_pos_accuracy": 0.560546875,
          "paraphrase_len4_program_exact": 0.0234375,
          "paraphrase_len8_direct_accuracy": NaN,
          "paraphrase_len8_direct_target_mass": NaN,
          "paraphrase_len8_executor_accuracy": 0.01171875,
          "paraphrase_len8_executor_target_mass": 0.010036191437393427,
          "paraphrase_len8_init_accuracy": 0.74609375,
          "paraphrase_len8_init_pos_accuracy": 0.921875,
          "paraphrase_len8_op_accuracy": 0.5654296875,
          "paraphrase_len8_arg_accuracy": 0.36767578125,
          "paraphrase_len8_op_pos_accuracy": 0.375,
          "paraphrase_len8_arg_pos_accuracy": 0.36328125,
          "paraphrase_len8_program_exact": 0.0,
          "paraphrase_len12_direct_accuracy": NaN,
          "paraphrase_len12_direct_target_mass": NaN,
          "paraphrase_len12_executor_accuracy": 0.01171875,
          "paraphrase_len12_executor_target_mass": 0.010314361657947302,
          "paraphrase_len12_init_accuracy": 0.7421875,
          "paraphrase_len12_init_pos_accuracy": 0.921875,
          "paraphrase_len12_op_accuracy": 0.5162760416666666,
          "paraphrase_len12_arg_accuracy": 0.2916666666666667,
          "paraphrase_len12_op_pos_accuracy": 0.2672526041666667,
          "paraphrase_len12_arg_pos_accuracy": 0.2708333333333333,
          "paraphrase_len12_program_exact": 0.0,
          "paraphrase_len24_direct_accuracy": NaN,
          "paraphrase_len24_direct_target_mass": NaN,
          "paraphrase_len24_executor_accuracy": 0.01953125,
          "paraphrase_len24_executor_target_mass": 0.01030915230512619,
          "paraphrase_len24_init_accuracy": 0.71484375,
          "paraphrase_len24_init_pos_accuracy": 0.921875,
          "paraphrase_len24_op_accuracy": 0.4361979166666667,
          "paraphrase_len24_arg_accuracy": 0.1640625,
          "paraphrase_len24_op_pos_accuracy": 0.12825520833333334,
          "paraphrase_len24_arg_pos_accuracy": 0.13297526041666666,
          "paraphrase_len24_program_exact": 0.0
        },
        {
          "stage": "trace_bootstrap",
          "local_step": 1600,
          "step": 1600,
          "loss": 8.43596076965332,
          "trace_loss_active": true,
          "lr_scale": 1.0,
          "executor_loss": 3.0295944213867188,
          "init_loss": 6.301762186922133e-05,
          "op_loss": 0.2963768541812897,
          "arg_loss": 0.5254144072532654,
          "init_attn_loss": 0.00018633034778758883,
          "op_attn_loss": 0.5555647015571594,
          "arg_attn_loss": 0.5689468383789062,
          "role_loss": 0.17654094099998474,
          "role_count_loss": 0.20004932582378387,
          "standard_len4_direct_accuracy": NaN,
          "standard_len4_direct_target_mass": NaN,
          "standard_len4_executor_accuracy": 0.73828125,
          "standard_len4_executor_target_mass": 0.08165208995342255,
          "standard_len4_init_accuracy": 1.0,
          "standard_len4_init_pos_accuracy": 1.0,
          "standard_len4_op_accuracy": 0.994140625,
          "standard_len4_arg_accuracy": 0.9326171875,
          "standard_len4_op_pos_accuracy": 1.0,
          "standard_len4_arg_pos_accuracy": 0.94140625,
          "standard_len4_program_exact": 0.73828125,
          "standard_len8_direct_accuracy": NaN,
          "standard_len8_direct_target_mass": NaN,
          "standard_len8_executor_accuracy": 0.51171875,
          "standard_len8_executor_target_mass": 0.01230234932154417,
          "standard_len8_init_accuracy": 1.0,
          "standard_len8_init_pos_accuracy": 1.0,
          "standard_len8_op_accuracy": 0.974609375,
          "standard_len8_arg_accuracy": 0.93798828125,
          "standard_len8_op_pos_accuracy": 1.0,
          "standard_len8_arg_pos_accuracy": 0.94580078125,
          "standard_len8_program_exact": 0.5078125,
          "standard_len12_direct_accuracy": NaN,
          "standard_len12_direct_target_mass": NaN,
          "standard_len12_executor_accuracy": 0.2421875,
          "standard_len12_executor_target_mass": 0.010357526130974293,
          "standard_len12_init_accuracy": 1.0,
          "standard_len12_init_pos_accuracy": 1.0,
          "standard_len12_op_accuracy": 0.9680989583333334,
          "standard_len12_arg_accuracy": 0.9114583333333334,
          "standard_len12_op_pos_accuracy": 0.9990234375,
          "standard_len12_arg_pos_accuracy": 0.9283854166666666,
          "standard_len12_program_exact": 0.23828125,
          "standard_len24_direct_accuracy": NaN,
          "standard_len24_direct_target_mass": NaN,
          "standard_len24_executor_accuracy": 0.01171875,
          "standard_len24_executor_target_mass": 0.010309250093996525,
          "standard_len24_init_accuracy": 1.0,
          "standard_len24_init_pos_accuracy": 1.0,
          "standard_len24_op_accuracy": 0.7403971354166666,
          "standard_len24_arg_accuracy": 0.6017252604166666,
          "standard_len24_op_pos_accuracy": 0.61376953125,
          "standard_len24_arg_pos_accuracy": 0.5831705729166666,
          "standard_len24_program_exact": 0.0,
          "paraphrase_len4_direct_accuracy": NaN,
          "paraphrase_len4_direct_target_mass": NaN,
          "paraphrase_len4_executor_accuracy": 0.0234375,
          "paraphrase_len4_executor_target_mass": 0.01477328222244978,
          "paraphrase_len4_init_accuracy": 0.55078125,
          "paraphrase_len4_init_pos_accuracy": 0.8359375,
          "paraphrase_len4_op_accuracy": 0.6640625,
          "paraphrase_len4_arg_accuracy": 0.541015625,
          "paraphrase_len4_op_pos_accuracy": 0.5927734375,
          "paraphrase_len4_arg_pos_accuracy": 0.57421875,
          "paraphrase_len4_program_exact": 0.015625,
          "paraphrase_len8_direct_accuracy": NaN,
          "paraphrase_len8_direct_target_mass": NaN,
          "paraphrase_len8_executor_accuracy": 0.00390625,
          "paraphrase_len8_executor_target_mass": 0.010181421414017677,
          "paraphrase_len8_init_accuracy": 0.53515625,
          "paraphrase_len8_init_pos_accuracy": 0.8203125,
          "paraphrase_len8_op_accuracy": 0.580078125,
          "paraphrase_len8_arg_accuracy": 0.3896484375,
          "paraphrase_len8_op_pos_accuracy": 0.41162109375,
          "paraphrase_len8_arg_pos_accuracy": 0.39599609375,
          "paraphrase_len8_program_exact": 0.0,
          "paraphrase_len12_direct_accuracy": NaN,
          "paraphrase_len12_direct_target_mass": NaN,
          "paraphrase_len12_executor_accuracy": 0.0078125,
          "paraphrase_len12_executor_target_mass": 0.010312501806765795,
          "paraphrase_len12_init_accuracy": 0.53125,
          "paraphrase_len12_init_pos_accuracy": 0.8515625,
          "paraphrase_len12_op_accuracy": 0.5211588541666666,
          "paraphrase_len12_arg_accuracy": 0.3069661458333333,
          "paraphrase_len12_op_pos_accuracy": 0.3043619791666667,
          "paraphrase_len12_arg_pos_accuracy": 0.3046875,
          "paraphrase_len12_program_exact": 0.0,
          "paraphrase_len24_direct_accuracy": NaN,
          "paraphrase_len24_direct_target_mass": NaN,
          "paraphrase_len24_executor_accuracy": 0.00390625,
          "paraphrase_len24_executor_target_mass": 0.010308923199772835,
          "paraphrase_len24_init_accuracy": 0.4921875,
          "paraphrase_len24_init_pos_accuracy": 0.85546875,
          "paraphrase_len24_op_accuracy": 0.4327799479166667,
          "paraphrase_len24_arg_accuracy": 0.17936197916666666,
          "paraphrase_len24_op_pos_accuracy": 0.15690104166666666,
          "paraphrase_len24_arg_pos_accuracy": 0.15576171875,
          "paraphrase_len24_program_exact": 0.0
        },
        {
          "stage": "answer_retention_low_lr",
          "local_step": 1,
          "step": 1601,
          "loss": 3.3286516666412354,
          "trace_loss_active": false,
          "lr_scale": 0.1,
          "executor_loss": 3.3286516666412354,
          "standard_len4_direct_accuracy": NaN,
          "standard_len4_direct_target_mass": NaN,
          "standard_len4_executor_accuracy": 0.73828125,
          "standard_len4_executor_target_mass": 0.08187716081738472,
          "standard_len4_init_accuracy": 1.0,
          "standard_len4_init_pos_accuracy": 1.0,
          "standard_len4_op_accuracy": 0.994140625,
          "standard_len4_arg_accuracy": 0.9326171875,
          "standard_len4_op_pos_accuracy": 1.0,
          "standard_len4_arg_pos_accuracy": 0.9423828125,
          "standard_len4_program_exact": 0.73828125,
          "standard_len8_direct_accuracy": NaN,
          "standard_len8_direct_target_mass": NaN,
          "standard_len8_executor_accuracy": 0.515625,
          "standard_len8_executor_target_mass": 0.012306882068514824,
          "standard_len8_init_accuracy": 1.0,
          "standard_len8_init_pos_accuracy": 1.0,
          "standard_len8_op_accuracy": 0.97607421875,
          "standard_len8_arg_accuracy": 0.93701171875,
          "standard_len8_op_pos_accuracy": 1.0,
          "standard_len8_arg_pos_accuracy": 0.9453125,
          "standard_len8_program_exact": 0.51171875,
          "standard_len12_direct_accuracy": NaN,
          "standard_len12_direct_target_mass": NaN,
          "standard_len12_executor_accuracy": 0.26171875,
          "standard_len12_executor_target_mass": 0.010358754079788923,
          "standard_len12_init_accuracy": 1.0,
          "standard_len12_init_pos_accuracy": 1.0,
          "standard_len12_op_accuracy": 0.9736328125,
          "standard_len12_arg_accuracy": 0.9104817708333334,
          "standard_len12_op_pos_accuracy": 0.9986979166666666,
          "standard_len12_arg_pos_accuracy": 0.9290364583333334,
          "standard_len12_program_exact": 0.2578125,
          "standard_len24_direct_accuracy": NaN,
          "standard_len24_direct_target_mass": NaN,
          "standard_len24_executor_accuracy": 0.015625,
          "standard_len24_executor_target_mass": 0.010309251956641674,
          "standard_len24_init_accuracy": 1.0,
          "standard_len24_init_pos_accuracy": 1.0,
          "standard_len24_op_accuracy": 0.7415364583333334,
          "standard_len24_arg_accuracy": 0.59765625,
          "standard_len24_op_pos_accuracy": 0.6095377604166666,
          "standard_len24_arg_pos_accuracy": 0.5789388020833334,
          "standard_len24_program_exact": 0.0,
          "paraphrase_len4_direct_accuracy": NaN,
          "paraphrase_len4_direct_target_mass": NaN,
          "paraphrase_len4_executor_accuracy": 0.0234375,
          "paraphrase_len4_executor_target_mass": 0.014789871871471405,
          "paraphrase_len4_init_accuracy": 0.546875,
          "paraphrase_len4_init_pos_accuracy": 0.8359375,
          "paraphrase_len4_op_accuracy": 0.6630859375,
          "paraphrase_len4_arg_accuracy": 0.5419921875,
          "paraphrase_len4_op_pos_accuracy": 0.59375,
          "paraphrase_len4_arg_pos_accuracy": 0.57421875,
          "paraphrase_len4_program_exact": 0.015625,
          "paraphrase_len8_direct_accuracy": NaN,
          "paraphrase_len8_direct_target_mass": NaN,
          "paraphrase_len8_executor_accuracy": 0.00390625,
          "paraphrase_len8_executor_target_mass": 0.010174406226724386,
          "paraphrase_len8_init_accuracy": 0.52734375,
          "paraphrase_len8_init_pos_accuracy": 0.81640625,
          "paraphrase_len8_op_accuracy": 0.57958984375,
          "paraphrase_len8_arg_accuracy": 0.39111328125,
          "paraphrase_len8_op_pos_accuracy": 0.41357421875,
          "paraphrase_len8_arg_pos_accuracy": 0.396484375,
          "paraphrase_len8_program_exact": 0.0,
          "paraphrase_len12_direct_accuracy": NaN,
          "paraphrase_len12_direct_target_mass": NaN,
          "paraphrase_len12_executor_accuracy": 0.0078125,
          "paraphrase_len12_executor_target_mass": 0.010313691105693579,
          "paraphrase_len12_init_accuracy": 0.53125,
          "paraphrase_len12_init_pos_accuracy": 0.8515625,
          "paraphrase_len12_op_accuracy": 0.5224609375,
          "paraphrase_len12_arg_accuracy": 0.3079427083333333,
          "paraphrase_len12_op_pos_accuracy": 0.3056640625,
          "paraphrase_len12_arg_pos_accuracy": 0.3063151041666667,
          "paraphrase_len12_program_exact": 0.0,
          "paraphrase_len24_direct_accuracy": NaN,
          "paraphrase_len24_direct_target_mass": NaN,
          "paraphrase_len24_executor_accuracy": 0.00390625,
          "paraphrase_len24_executor_target_mass": 0.01030893437564373,
          "paraphrase_len24_init_accuracy": 0.4921875,
          "paraphrase_len24_init_pos_accuracy": 0.85546875,
          "paraphrase_len24_op_accuracy": 0.4327799479166667,
          "paraphrase_len24_arg_accuracy": 0.1806640625,
          "paraphrase_len24_op_pos_accuracy": 0.15738932291666666,
          "paraphrase_len24_arg_pos_accuracy": 0.15738932291666666,
          "paraphrase_len24_program_exact": 0.0
        },
        {
          "stage": "answer_retention_low_lr",
          "local_step": 800,
          "step": 2400,
          "loss": 3.012664794921875,
          "trace_loss_active": false,
          "lr_scale": 0.1,
          "executor_loss": 3.012664794921875,
          "standard_len4_direct_accuracy": NaN,
          "standard_len4_direct_target_mass": NaN,
          "standard_len4_executor_accuracy": 0.78125,
          "standard_len4_executor_target_mass": 0.16868610680103302,
          "standard_len4_init_accuracy": 1.0,
          "standard_len4_init_pos_accuracy": 1.0,
          "standard_len4_op_accuracy": 0.97265625,
          "standard_len4_arg_accuracy": 0.9638671875,
          "standard_len4_op_pos_accuracy": 0.478515625,
          "standard_len4_arg_pos_accuracy": 0.50390625,
          "standard_len4_program_exact": 0.78125,
          "standard_len8_direct_accuracy": NaN,
          "standard_len8_direct_target_mass": NaN,
          "standard_len8_executor_accuracy": 0.06640625,
          "standard_len8_executor_target_mass": 0.01282868953421712,
          "standard_len8_init_accuracy": 1.0,
          "standard_len8_init_pos_accuracy": 1.0,
          "standard_len8_op_accuracy": 0.869140625,
          "standard_len8_arg_accuracy": 0.81591796875,
          "standard_len8_op_pos_accuracy": 0.2421875,
          "standard_len8_arg_pos_accuracy": 0.29150390625,
          "standard_len8_program_exact": 0.0546875,
          "standard_len12_direct_accuracy": NaN,
          "standard_len12_direct_target_mass": NaN,
          "standard_len12_executor_accuracy": 0.0078125,
          "standard_len12_executor_target_mass": 0.01031169481575489,
          "standard_len12_init_accuracy": 1.0,
          "standard_len12_init_pos_accuracy": 1.0,
          "standard_len12_op_accuracy": 0.76953125,
          "standard_len12_arg_accuracy": 0.669921875,
          "standard_len12_op_pos_accuracy": 0.16015625,
          "standard_len12_arg_pos_accuracy": 0.19954427083333334,
          "standard_len12_program_exact": 0.0,
          "standard_len24_direct_accuracy": NaN,
          "standard_len24_direct_target_mass": NaN,
          "standard_len24_executor_accuracy": 0.0,
          "standard_len24_executor_target_mass": 0.010309291537851095,
          "standard_len24_init_accuracy": 1.0,
          "standard_len24_init_pos_accuracy": 1.0,
          "standard_len24_op_accuracy": 0.6236979166666666,
          "standard_len24_arg_accuracy": 0.43701171875,
          "standard_len24_op_pos_accuracy": 0.08121744791666667,
          "standard_len24_arg_pos_accuracy": 0.10074869791666667,
          "standard_len24_program_exact": 0.0,
          "paraphrase_len4_direct_accuracy": NaN,
          "paraphrase_len4_direct_target_mass": NaN,
          "paraphrase_len4_executor_accuracy": 0.01953125,
          "paraphrase_len4_executor_target_mass": 0.014725316315889359,
          "paraphrase_len4_init_accuracy": 0.59765625,
          "paraphrase_len4_init_pos_accuracy": 0.8203125,
          "paraphrase_len4_op_accuracy": 0.62109375,
          "paraphrase_len4_arg_accuracy": 0.5341796875,
          "paraphrase_len4_op_pos_accuracy": 0.1806640625,
          "paraphrase_len4_arg_pos_accuracy": 0.404296875,
          "paraphrase_len4_program_exact": 0.0078125,
          "paraphrase_len8_direct_accuracy": NaN,
          "paraphrase_len8_direct_target_mass": NaN,
          "paraphrase_len8_executor_accuracy": 0.0078125,
          "paraphrase_len8_executor_target_mass": 0.00983346113935113,
          "paraphrase_len8_init_accuracy": 0.5703125,
          "paraphrase_len8_init_pos_accuracy": 0.8125,
          "paraphrase_len8_op_accuracy": 0.56591796875,
          "paraphrase_len8_arg_accuracy": 0.40234375,
          "paraphrase_len8_op_pos_accuracy": 0.19140625,
          "paraphrase_len8_arg_pos_accuracy": 0.29345703125,
          "paraphrase_len8_program_exact": 0.0,
          "paraphrase_len12_direct_accuracy": NaN,
          "paraphrase_len12_direct_target_mass": NaN,
          "paraphrase_len12_executor_accuracy": 0.00390625,
          "paraphrase_len12_executor_target_mass": 0.010364487767219543,
          "paraphrase_len12_init_accuracy": 0.5859375,
          "paraphrase_len12_init_pos_accuracy": 0.796875,
          "paraphrase_len12_op_accuracy": 0.544921875,
          "paraphrase_len12_arg_accuracy": 0.3600260416666667,
          "paraphrase_len12_op_pos_accuracy": 0.1953125,
          "paraphrase_len12_arg_pos_accuracy": 0.2760416666666667,
          "paraphrase_len12_program_exact": 0.0,
          "paraphrase_len24_direct_accuracy": NaN,
          "paraphrase_len24_direct_target_mass": NaN,
          "paraphrase_len24_executor_accuracy": 0.01171875,
          "paraphrase_len24_executor_target_mass": 0.010307228192687035,
          "paraphrase_len24_init_accuracy": 0.55078125,
          "paraphrase_len24_init_pos_accuracy": 0.8203125,
          "paraphrase_len24_op_accuracy": 0.5135091145833334,
          "paraphrase_len24_arg_accuracy": 0.31494140625,
          "paraphrase_len24_op_pos_accuracy": 0.22200520833333334,
          "paraphrase_len24_arg_pos_accuracy": 0.2501627604166667,
          "paraphrase_len24_program_exact": 0.0
        }
      ],
      "final_metrics": {
        "standard_len4": {
          "variant": "compiler_trace_then_answer_low_lr",
          "n": 256.0,
          "direct_accuracy": NaN,
          "direct_target_mass": NaN,
          "executor_accuracy": 0.78125,
          "executor_target_mass": 0.16868610680103302,
          "init_accuracy": 1.0,
          "init_pos_accuracy": 1.0,
          "op_accuracy": 0.97265625,
          "arg_accuracy": 0.9638671875,
          "op_pos_accuracy": 0.478515625,
          "arg_pos_accuracy": 0.50390625,
          "program_exact": 0.78125
        },
        "standard_len8": {
          "variant": "compiler_trace_then_answer_low_lr",
          "n": 256.0,
          "direct_accuracy": NaN,
          "direct_target_mass": NaN,
          "executor_accuracy": 0.06640625,
          "executor_target_mass": 0.01282868953421712,
          "init_accuracy": 1.0,
          "init_pos_accuracy": 1.0,
          "op_accuracy": 0.869140625,
          "arg_accuracy": 0.81591796875,
          "op_pos_accuracy": 0.2421875,
          "arg_pos_accuracy": 0.29150390625,
          "program_exact": 0.0546875
        },
        "standard_len12": {
          "variant": "compiler_trace_then_answer_low_lr",
          "n": 256.0,
          "direct_accuracy": NaN,
          "direct_target_mass": NaN,
          "executor_accuracy": 0.0078125,
          "executor_target_mass": 0.01031169481575489,
          "init_accuracy": 1.0,
          "init_pos_accuracy": 1.0,
          "op_accuracy": 0.76953125,
          "arg_accuracy": 0.669921875,
          "op_pos_accuracy": 0.16015625,
          "arg_pos_accuracy": 0.19954427083333334,
          "program_exact": 0.0
        },
        "standard_len24": {
          "variant": "compiler_trace_then_answer_low_lr",
          "n": 256.0,
          "direct_accuracy": NaN,
          "direct_target_mass": NaN,
          "executor_accuracy": 0.0,
          "executor_target_mass": 0.010309291537851095,
          "init_accuracy": 1.0,
          "init_pos_accuracy": 1.0,
          "op_accuracy": 0.6236979166666666,
          "arg_accuracy": 0.43701171875,
          "op_pos_accuracy": 0.08121744791666667,
          "arg_pos_accuracy": 0.10074869791666667,
          "program_exact": 0.0
        },
        "paraphrase_len4": {
          "variant": "compiler_trace_then_answer_low_lr",
          "n": 256.0,
          "direct_accuracy": NaN,
          "direct_target_mass": NaN,
          "executor_accuracy": 0.01953125,
          "executor_target_mass": 0.014725316315889359,
          "init_accuracy": 0.59765625,
          "init_pos_accuracy": 0.8203125,
          "op_accuracy": 0.62109375,
          "arg_accuracy": 0.5341796875,
          "op_pos_accuracy": 0.1806640625,
          "arg_pos_accuracy": 0.404296875,
          "program_exact": 0.0078125
        },
        "paraphrase_len8": {
          "variant": "compiler_trace_then_answer_low_lr",
          "n": 256.0,
          "direct_accuracy": NaN,
          "direct_target_mass": NaN,
          "executor_accuracy": 0.0078125,
          "executor_target_mass": 0.00983346113935113,
          "init_accuracy": 0.5703125,
          "init_pos_accuracy": 0.8125,
          "op_accuracy": 0.56591796875,
          "arg_accuracy": 0.40234375,
          "op_pos_accuracy": 0.19140625,
          "arg_pos_accuracy": 0.29345703125,
          "program_exact": 0.0
        },
        "paraphrase_len12": {
          "variant": "compiler_trace_then_answer_low_lr",
          "n": 256.0,
          "direct_accuracy": NaN,
          "direct_target_mass": NaN,
          "executor_accuracy": 0.00390625,
          "executor_target_mass": 0.010364487767219543,
          "init_accuracy": 0.5859375,
          "init_pos_accuracy": 0.796875,
          "op_accuracy": 0.544921875,
          "arg_accuracy": 0.3600260416666667,
          "op_pos_accuracy": 0.1953125,
          "arg_pos_accuracy": 0.2760416666666667,
          "program_exact": 0.0
        },
        "paraphrase_len24": {
          "variant": "compiler_trace_then_answer_low_lr",
          "n": 256.0,
          "direct_accuracy": NaN,
          "direct_target_mass": NaN,
          "executor_accuracy": 0.01171875,
          "executor_target_mass": 0.010307228192687035,
          "init_accuracy": 0.55078125,
          "init_pos_accuracy": 0.8203125,
          "op_accuracy": 0.5135091145833334,
          "arg_accuracy": 0.31494140625,
          "op_pos_accuracy": 0.22200520833333334,
          "arg_pos_accuracy": 0.2501627604166667,
          "program_exact": 0.0
        }
      }
    }
  }
}