{
  "args": {
    "model_id": "hf-internal-testing/tiny-random-LlamaForCausalLM",
    "text_model_fallback": "Qwen/Qwen3-4B",
    "model_class": "causal",
    "load_in_4bit": false,
    "torch_dtype": "fp32",
    "device_map": "auto",
    "modulus": 17,
    "max_steps": 6,
    "train_min_len": 1,
    "train_max_len": 2,
    "answer_train_min_len": 1,
    "answer_train_max_len": 4,
    "eval_lengths": "2,4",
    "train_template_mode": "standard",
    "answer_train_template_mode": "standard",
    "eval_template_modes": "standard,paraphrase",
    "train_size": 24,
    "answer_train_size": 24,
    "eval_size": 12,
    "feature_batch_size": 8,
    "train_batch_size": 8,
    "eval_batch_size": 16,
    "bootstrap_steps": 4,
    "answer_steps": 4,
    "eval_every": 4,
    "stage_eval_every": 2,
    "lr": 0.001,
    "answer_lr_scale": 0.1,
    "weight_decay": 0.01,
    "grad_clip": 1.0,
    "head_width": 64,
    "trace_loss_weight": 1.0,
    "executor_loss_weight": 1.0,
    "variants": "compiler_trace,compiler_trace_then_answer",
    "max_length": 512,
    "feature_seq_len": 128,
    "seed": 7,
    "output_dir": "experiments/qwen_span_free_compiler/runs/smoke_tiny_token_pool",
    "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": "AutoModelForCausalLM",
    "peft_installed": true,
    "gpu_name": "NVIDIA RTX 6000 Ada Generation",
    "gpu_vram_gb": 47.383
  },
  "hidden_dim": 16,
  "dataset": {
    "bootstrap_train_size": 24,
    "answer_train_size": 24,
    "eval_size": 12,
    "bootstrap_train_lengths": [
      1,
      2
    ],
    "answer_train_lengths": [
      1,
      4
    ],
    "eval_lengths": [
      2,
      4
    ],
    "train_template_mode": "standard",
    "answer_train_template_mode": "standard",
    "eval_template_modes": [
      "standard",
      "paraphrase"
    ],
    "feature_seq_len": 128
  },
  "variants": {
    "compiler_trace": {
      "variant": "compiler_trace",
      "train_seconds": 1.1316850185394287,
      "checkpoints": [
        "large_artifacts/qwen_span_free_compiler/checkpoints/smoke_tiny_token_pool/compiler_trace/checkpoint_final.pt"
      ],
      "train_log": [
        {
          "stage": "trace_bootstrap",
          "local_step": 1,
          "step": 1,
          "loss": 9.748407363891602,
          "trace_loss_active": true,
          "lr_scale": 1.0,
          "executor_loss": 2.841398239135742,
          "init_loss": 2.9160172939300537,
          "op_loss": 1.0656849145889282,
          "arg_loss": 2.925307273864746,
          "standard_len2_direct_accuracy": NaN,
          "standard_len2_direct_target_mass": NaN,
          "standard_len2_executor_accuracy": 0.08333333333333333,
          "standard_len2_executor_target_mass": 0.06059951583544413,
          "standard_len2_init_accuracy": 0.0,
          "standard_len2_op_accuracy": 0.25,
          "standard_len2_arg_accuracy": 0.041666666666666664,
          "standard_len2_program_exact": 0.0,
          "standard_len4_direct_accuracy": NaN,
          "standard_len4_direct_target_mass": NaN,
          "standard_len4_executor_accuracy": 0.08333333333333333,
          "standard_len4_executor_target_mass": 0.057822595040003456,
          "standard_len4_init_accuracy": 0.0,
          "standard_len4_op_accuracy": 0.2916666666666667,
          "standard_len4_arg_accuracy": 0.0625,
          "standard_len4_program_exact": 0.0,
          "paraphrase_len2_direct_accuracy": NaN,
          "paraphrase_len2_direct_target_mass": NaN,
          "paraphrase_len2_executor_accuracy": 0.0,
          "paraphrase_len2_executor_target_mass": 0.05911166469256083,
          "paraphrase_len2_init_accuracy": 0.0,
          "paraphrase_len2_op_accuracy": 0.25,
          "paraphrase_len2_arg_accuracy": 0.0,
          "paraphrase_len2_program_exact": 0.0,
          "paraphrase_len4_direct_accuracy": NaN,
          "paraphrase_len4_direct_target_mass": NaN,
          "paraphrase_len4_executor_accuracy": 0.0,
          "paraphrase_len4_executor_target_mass": 0.059172396858533226,
          "paraphrase_len4_init_accuracy": 0.0,
          "paraphrase_len4_op_accuracy": 0.2916666666666667,
          "paraphrase_len4_arg_accuracy": 0.0625,
          "paraphrase_len4_program_exact": 0.0
        },
        {
          "stage": "trace_bootstrap",
          "local_step": 2,
          "step": 2,
          "loss": 9.628660202026367,
          "trace_loss_active": true,
          "lr_scale": 1.0,
          "executor_loss": 2.848123073577881,
          "init_loss": 2.830629825592041,
          "op_loss": 1.1004564762115479,
          "arg_loss": 2.84945011138916,
          "standard_len2_direct_accuracy": NaN,
          "standard_len2_direct_target_mass": NaN,
          "standard_len2_executor_accuracy": 0.0,
          "standard_len2_executor_target_mass": 0.0606837123632431,
          "standard_len2_init_accuracy": 0.16666666666666666,
          "standard_len2_op_accuracy": 0.25,
          "standard_len2_arg_accuracy": 0.041666666666666664,
          "standard_len2_program_exact": 0.0,
          "standard_len4_direct_accuracy": NaN,
          "standard_len4_direct_target_mass": NaN,
          "standard_len4_executor_accuracy": 0.08333333333333333,
          "standard_len4_executor_target_mass": 0.05776595572630564,
          "standard_len4_init_accuracy": 0.16666666666666666,
          "standard_len4_op_accuracy": 0.2916666666666667,
          "standard_len4_arg_accuracy": 0.08333333333333333,
          "standard_len4_program_exact": 0.0,
          "paraphrase_len2_direct_accuracy": NaN,
          "paraphrase_len2_direct_target_mass": NaN,
          "paraphrase_len2_executor_accuracy": 0.08333333333333333,
          "paraphrase_len2_executor_target_mass": 0.05912197629610697,
          "paraphrase_len2_init_accuracy": 0.0,
          "paraphrase_len2_op_accuracy": 0.3333333333333333,
          "paraphrase_len2_arg_accuracy": 0.041666666666666664,
          "paraphrase_len2_program_exact": 0.0,
          "paraphrase_len4_direct_accuracy": NaN,
          "paraphrase_len4_direct_target_mass": NaN,
          "paraphrase_len4_executor_accuracy": 0.16666666666666666,
          "paraphrase_len4_executor_target_mass": 0.059181337555249534,
          "paraphrase_len4_init_accuracy": 0.0,
          "paraphrase_len4_op_accuracy": 0.2916666666666667,
          "paraphrase_len4_arg_accuracy": 0.0625,
          "paraphrase_len4_program_exact": 0.0
        },
        {
          "stage": "trace_bootstrap",
          "local_step": 4,
          "step": 4,
          "loss": 9.512877464294434,
          "trace_loss_active": true,
          "lr_scale": 1.0,
          "executor_loss": 2.8117942810058594,
          "init_loss": 2.8894765377044678,
          "op_loss": 1.027420163154602,
          "arg_loss": 2.784186601638794,
          "standard_len2_direct_accuracy": NaN,
          "standard_len2_direct_target_mass": NaN,
          "standard_len2_executor_accuracy": 0.0,
          "standard_len2_executor_target_mass": 0.06083860993385315,
          "standard_len2_init_accuracy": 0.16666666666666666,
          "standard_len2_op_accuracy": 0.4166666666666667,
          "standard_len2_arg_accuracy": 0.16666666666666666,
          "standard_len2_program_exact": 0.0,
          "standard_len4_direct_accuracy": NaN,
          "standard_len4_direct_target_mass": NaN,
          "standard_len4_executor_accuracy": 0.08333333333333333,
          "standard_len4_executor_target_mass": 0.05764159063498179,
          "standard_len4_init_accuracy": 0.16666666666666666,
          "standard_len4_op_accuracy": 0.2916666666666667,
          "standard_len4_arg_accuracy": 0.08333333333333333,
          "standard_len4_program_exact": 0.0,
          "paraphrase_len2_direct_accuracy": NaN,
          "paraphrase_len2_direct_target_mass": NaN,
          "paraphrase_len2_executor_accuracy": 0.08333333333333333,
          "paraphrase_len2_executor_target_mass": 0.059142351150512695,
          "paraphrase_len2_init_accuracy": 0.0,
          "paraphrase_len2_op_accuracy": 0.375,
          "paraphrase_len2_arg_accuracy": 0.041666666666666664,
          "paraphrase_len2_program_exact": 0.0,
          "paraphrase_len4_direct_accuracy": NaN,
          "paraphrase_len4_direct_target_mass": NaN,
          "paraphrase_len4_executor_accuracy": 0.08333333333333333,
          "paraphrase_len4_executor_target_mass": 0.059219539165496826,
          "paraphrase_len4_init_accuracy": 0.08333333333333333,
          "paraphrase_len4_op_accuracy": 0.3541666666666667,
          "paraphrase_len4_arg_accuracy": 0.020833333333333332,
          "paraphrase_len4_program_exact": 0.0
        },
        {
          "stage": "trace_continuation",
          "local_step": 1,
          "step": 5,
          "loss": 9.93716812133789,
          "trace_loss_active": true,
          "lr_scale": 1.0,
          "executor_loss": 2.851499080657959,
          "init_loss": 2.9231982231140137,
          "op_loss": 1.178879976272583,
          "arg_loss": 2.9835901260375977,
          "standard_len2_direct_accuracy": NaN,
          "standard_len2_direct_target_mass": NaN,
          "standard_len2_executor_accuracy": 0.0,
          "standard_len2_executor_target_mass": 0.06093661983807882,
          "standard_len2_init_accuracy": 0.16666666666666666,
          "standard_len2_op_accuracy": 0.3333333333333333,
          "standard_len2_arg_accuracy": 0.16666666666666666,
          "standard_len2_program_exact": 0.0,
          "standard_len4_direct_accuracy": NaN,
          "standard_len4_direct_target_mass": NaN,
          "standard_len4_executor_accuracy": 0.0,
          "standard_len4_executor_target_mass": 0.0575664738814036,
          "standard_len4_init_accuracy": 0.16666666666666666,
          "standard_len4_op_accuracy": 0.2916666666666667,
          "standard_len4_arg_accuracy": 0.08333333333333333,
          "standard_len4_program_exact": 0.0,
          "paraphrase_len2_direct_accuracy": NaN,
          "paraphrase_len2_direct_target_mass": NaN,
          "paraphrase_len2_executor_accuracy": 0.08333333333333333,
          "paraphrase_len2_executor_target_mass": 0.0591505765914917,
          "paraphrase_len2_init_accuracy": 0.0,
          "paraphrase_len2_op_accuracy": 0.25,
          "paraphrase_len2_arg_accuracy": 0.041666666666666664,
          "paraphrase_len2_program_exact": 0.0,
          "paraphrase_len4_direct_accuracy": NaN,
          "paraphrase_len4_direct_target_mass": NaN,
          "paraphrase_len4_executor_accuracy": 0.08333333333333333,
          "paraphrase_len4_executor_target_mass": 0.05925719936688741,
          "paraphrase_len4_init_accuracy": 0.08333333333333333,
          "paraphrase_len4_op_accuracy": 0.4375,
          "paraphrase_len4_arg_accuracy": 0.020833333333333332,
          "paraphrase_len4_program_exact": 0.0
        },
        {
          "stage": "trace_continuation",
          "local_step": 2,
          "step": 6,
          "loss": 9.921004295349121,
          "trace_loss_active": true,
          "lr_scale": 1.0,
          "executor_loss": 2.809476613998413,
          "init_loss": 2.904829978942871,
          "op_loss": 1.1932073831558228,
          "arg_loss": 3.0134899616241455,
          "standard_len2_direct_accuracy": NaN,
          "standard_len2_direct_target_mass": NaN,
          "standard_len2_executor_accuracy": 0.0,
          "standard_len2_executor_target_mass": 0.061002482970555626,
          "standard_len2_init_accuracy": 0.16666666666666666,
          "standard_len2_op_accuracy": 0.3333333333333333,
          "standard_len2_arg_accuracy": 0.16666666666666666,
          "standard_len2_program_exact": 0.0,
          "standard_len4_direct_accuracy": NaN,
          "standard_len4_direct_target_mass": NaN,
          "standard_len4_executor_accuracy": 0.0,
          "standard_len4_executor_target_mass": 0.05750646193822225,
          "standard_len4_init_accuracy": 0.16666666666666666,
          "standard_len4_op_accuracy": 0.2916666666666667,
          "standard_len4_arg_accuracy": 0.08333333333333333,
          "standard_len4_program_exact": 0.0,
          "paraphrase_len2_direct_accuracy": NaN,
          "paraphrase_len2_direct_target_mass": NaN,
          "paraphrase_len2_executor_accuracy": 0.08333333333333333,
          "paraphrase_len2_executor_target_mass": 0.05915412803490957,
          "paraphrase_len2_init_accuracy": 0.0,
          "paraphrase_len2_op_accuracy": 0.2916666666666667,
          "paraphrase_len2_arg_accuracy": 0.041666666666666664,
          "paraphrase_len2_program_exact": 0.0,
          "paraphrase_len4_direct_accuracy": NaN,
          "paraphrase_len4_direct_target_mass": NaN,
          "paraphrase_len4_executor_accuracy": 0.08333333333333333,
          "paraphrase_len4_executor_target_mass": 0.059285337726275124,
          "paraphrase_len4_init_accuracy": 0.0,
          "paraphrase_len4_op_accuracy": 0.4166666666666667,
          "paraphrase_len4_arg_accuracy": 0.020833333333333332,
          "paraphrase_len4_program_exact": 0.0
        },
        {
          "stage": "trace_continuation",
          "local_step": 4,
          "step": 8,
          "loss": 9.709188461303711,
          "trace_loss_active": true,
          "lr_scale": 1.0,
          "executor_loss": 2.8034093379974365,
          "init_loss": 2.931992769241333,
          "op_loss": 1.1288580894470215,
          "arg_loss": 2.8449277877807617,
          "standard_len2_direct_accuracy": NaN,
          "standard_len2_direct_target_mass": NaN,
          "standard_len2_executor_accuracy": 0.0,
          "standard_len2_executor_target_mass": 0.060962398846944175,
          "standard_len2_init_accuracy": 0.16666666666666666,
          "standard_len2_op_accuracy": 0.4583333333333333,
          "standard_len2_arg_accuracy": 0.125,
          "standard_len2_program_exact": 0.0,
          "standard_len4_direct_accuracy": NaN,
          "standard_len4_direct_target_mass": NaN,
          "standard_len4_executor_accuracy": 0.0,
          "standard_len4_executor_target_mass": 0.057516664266586304,
          "standard_len4_init_accuracy": 0.16666666666666666,
          "standard_len4_op_accuracy": 0.2916666666666667,
          "standard_len4_arg_accuracy": 0.08333333333333333,
          "standard_len4_program_exact": 0.0,
          "paraphrase_len2_direct_accuracy": NaN,
          "paraphrase_len2_direct_target_mass": NaN,
          "paraphrase_len2_executor_accuracy": 0.16666666666666666,
          "paraphrase_len2_executor_target_mass": 0.059134066104888916,
          "paraphrase_len2_init_accuracy": 0.0,
          "paraphrase_len2_op_accuracy": 0.2916666666666667,
          "paraphrase_len2_arg_accuracy": 0.041666666666666664,
          "paraphrase_len2_program_exact": 0.0,
          "paraphrase_len4_direct_accuracy": NaN,
          "paraphrase_len4_direct_target_mass": NaN,
          "paraphrase_len4_executor_accuracy": 0.08333333333333333,
          "paraphrase_len4_executor_target_mass": 0.05929488937060038,
          "paraphrase_len4_init_accuracy": 0.0,
          "paraphrase_len4_op_accuracy": 0.4166666666666667,
          "paraphrase_len4_arg_accuracy": 0.0625,
          "paraphrase_len4_program_exact": 0.0
        }
      ],
      "final_metrics": {
        "standard_len2": {
          "variant": "compiler_trace",
          "n": 12.0,
          "direct_accuracy": NaN,
          "direct_target_mass": NaN,
          "executor_accuracy": 0.0,
          "executor_target_mass": 0.060962398846944175,
          "init_accuracy": 0.16666666666666666,
          "op_accuracy": 0.4583333333333333,
          "arg_accuracy": 0.125,
          "program_exact": 0.0
        },
        "standard_len4": {
          "variant": "compiler_trace",
          "n": 12.0,
          "direct_accuracy": NaN,
          "direct_target_mass": NaN,
          "executor_accuracy": 0.0,
          "executor_target_mass": 0.057516664266586304,
          "init_accuracy": 0.16666666666666666,
          "op_accuracy": 0.2916666666666667,
          "arg_accuracy": 0.08333333333333333,
          "program_exact": 0.0
        },
        "paraphrase_len2": {
          "variant": "compiler_trace",
          "n": 12.0,
          "direct_accuracy": NaN,
          "direct_target_mass": NaN,
          "executor_accuracy": 0.16666666666666666,
          "executor_target_mass": 0.059134066104888916,
          "init_accuracy": 0.0,
          "op_accuracy": 0.2916666666666667,
          "arg_accuracy": 0.041666666666666664,
          "program_exact": 0.0
        },
        "paraphrase_len4": {
          "variant": "compiler_trace",
          "n": 12.0,
          "direct_accuracy": NaN,
          "direct_target_mass": NaN,
          "executor_accuracy": 0.08333333333333333,
          "executor_target_mass": 0.05929488937060038,
          "init_accuracy": 0.0,
          "op_accuracy": 0.4166666666666667,
          "arg_accuracy": 0.0625,
          "program_exact": 0.0
        }
      }
    },
    "compiler_trace_then_answer": {
      "variant": "compiler_trace_then_answer",
      "train_seconds": 0.8979711532592773,
      "checkpoints": [
        "large_artifacts/qwen_span_free_compiler/checkpoints/smoke_tiny_token_pool/compiler_trace_then_answer/checkpoint_final.pt"
      ],
      "train_log": [
        {
          "stage": "trace_bootstrap",
          "local_step": 1,
          "step": 1,
          "loss": 9.634632110595703,
          "trace_loss_active": true,
          "lr_scale": 1.0,
          "executor_loss": 2.859280824661255,
          "init_loss": 2.867441177368164,
          "op_loss": 1.1015413999557495,
          "arg_loss": 2.8063690662384033,
          "standard_len2_direct_accuracy": NaN,
          "standard_len2_direct_target_mass": NaN,
          "standard_len2_executor_accuracy": 0.08333333333333333,
          "standard_len2_executor_target_mass": 0.06179186701774597,
          "standard_len2_init_accuracy": 0.08333333333333333,
          "standard_len2_op_accuracy": 0.4583333333333333,
          "standard_len2_arg_accuracy": 0.125,
          "standard_len2_program_exact": 0.0,
          "standard_len4_direct_accuracy": NaN,
          "standard_len4_direct_target_mass": NaN,
          "standard_len4_executor_accuracy": 0.16666666666666666,
          "standard_len4_executor_target_mass": 0.05686754981676737,
          "standard_len4_init_accuracy": 0.0,
          "standard_len4_op_accuracy": 0.2916666666666667,
          "standard_len4_arg_accuracy": 0.0625,
          "standard_len4_program_exact": 0.0,
          "paraphrase_len2_direct_accuracy": NaN,
          "paraphrase_len2_direct_target_mass": NaN,
          "paraphrase_len2_executor_accuracy": 0.0,
          "paraphrase_len2_executor_target_mass": 0.05960864325364431,
          "paraphrase_len2_init_accuracy": 0.16666666666666666,
          "paraphrase_len2_op_accuracy": 0.3333333333333333,
          "paraphrase_len2_arg_accuracy": 0.20833333333333334,
          "paraphrase_len2_program_exact": 0.0,
          "paraphrase_len4_direct_accuracy": NaN,
          "paraphrase_len4_direct_target_mass": NaN,
          "paraphrase_len4_executor_accuracy": 0.08333333333333333,
          "paraphrase_len4_executor_target_mass": 0.05962463219960531,
          "paraphrase_len4_init_accuracy": 0.08333333333333333,
          "paraphrase_len4_op_accuracy": 0.3958333333333333,
          "paraphrase_len4_arg_accuracy": 0.020833333333333332,
          "paraphrase_len4_program_exact": 0.0
        },
        {
          "stage": "trace_bootstrap",
          "local_step": 2,
          "step": 2,
          "loss": 9.598407745361328,
          "trace_loss_active": true,
          "lr_scale": 1.0,
          "executor_loss": 2.8103137016296387,
          "init_loss": 2.8824219703674316,
          "op_loss": 1.054774284362793,
          "arg_loss": 2.8508970737457275,
          "standard_len2_direct_accuracy": NaN,
          "standard_len2_direct_target_mass": NaN,
          "standard_len2_executor_accuracy": 0.08333333333333333,
          "standard_len2_executor_target_mass": 0.06191416581471761,
          "standard_len2_init_accuracy": 0.08333333333333333,
          "standard_len2_op_accuracy": 0.4583333333333333,
          "standard_len2_arg_accuracy": 0.125,
          "standard_len2_program_exact": 0.0,
          "standard_len4_direct_accuracy": NaN,
          "standard_len4_direct_target_mass": NaN,
          "standard_len4_executor_accuracy": 0.16666666666666666,
          "standard_len4_executor_target_mass": 0.056760430335998535,
          "standard_len4_init_accuracy": 0.0,
          "standard_len4_op_accuracy": 0.2708333333333333,
          "standard_len4_arg_accuracy": 0.0625,
          "standard_len4_program_exact": 0.0,
          "paraphrase_len2_direct_accuracy": NaN,
          "paraphrase_len2_direct_target_mass": NaN,
          "paraphrase_len2_executor_accuracy": 0.0,
          "paraphrase_len2_executor_target_mass": 0.059609681367874146,
          "paraphrase_len2_init_accuracy": 0.0,
          "paraphrase_len2_op_accuracy": 0.3333333333333333,
          "paraphrase_len2_arg_accuracy": 0.20833333333333334,
          "paraphrase_len2_program_exact": 0.0,
          "paraphrase_len4_direct_accuracy": NaN,
          "paraphrase_len4_direct_target_mass": NaN,
          "paraphrase_len4_executor_accuracy": 0.16666666666666666,
          "paraphrase_len4_executor_target_mass": 0.059660752614339195,
          "paraphrase_len4_init_accuracy": 0.16666666666666666,
          "paraphrase_len4_op_accuracy": 0.3958333333333333,
          "paraphrase_len4_arg_accuracy": 0.020833333333333332,
          "paraphrase_len4_program_exact": 0.0
        },
        {
          "stage": "trace_bootstrap",
          "local_step": 4,
          "step": 4,
          "loss": 9.358291625976562,
          "trace_loss_active": true,
          "lr_scale": 1.0,
          "executor_loss": 2.8613486289978027,
          "init_loss": 2.8348188400268555,
          "op_loss": 0.9588616490364075,
          "arg_loss": 2.703263282775879,
          "standard_len2_direct_accuracy": NaN,
          "standard_len2_direct_target_mass": NaN,
          "standard_len2_executor_accuracy": 0.08333333333333333,
          "standard_len2_executor_target_mass": 0.061824992299079895,
          "standard_len2_init_accuracy": 0.08333333333333333,
          "standard_len2_op_accuracy": 0.4583333333333333,
          "standard_len2_arg_accuracy": 0.125,
          "standard_len2_program_exact": 0.0,
          "standard_len4_direct_accuracy": NaN,
          "standard_len4_direct_target_mass": NaN,
          "standard_len4_executor_accuracy": 0.0,
          "standard_len4_executor_target_mass": 0.056733076771100364,
          "standard_len4_init_accuracy": 0.0,
          "standard_len4_op_accuracy": 0.2916666666666667,
          "standard_len4_arg_accuracy": 0.08333333333333333,
          "standard_len4_program_exact": 0.0,
          "paraphrase_len2_direct_accuracy": NaN,
          "paraphrase_len2_direct_target_mass": NaN,
          "paraphrase_len2_executor_accuracy": 0.08333333333333333,
          "paraphrase_len2_executor_target_mass": 0.059524521231651306,
          "paraphrase_len2_init_accuracy": 0.0,
          "paraphrase_len2_op_accuracy": 0.3333333333333333,
          "paraphrase_len2_arg_accuracy": 0.16666666666666666,
          "paraphrase_len2_program_exact": 0.0,
          "paraphrase_len4_direct_accuracy": NaN,
          "paraphrase_len4_direct_target_mass": NaN,
          "paraphrase_len4_executor_accuracy": 0.16666666666666666,
          "paraphrase_len4_executor_target_mass": 0.059718549251556396,
          "paraphrase_len4_init_accuracy": 0.08333333333333333,
          "paraphrase_len4_op_accuracy": 0.3958333333333333,
          "paraphrase_len4_arg_accuracy": 0.0,
          "paraphrase_len4_program_exact": 0.0
        },
        {
          "stage": "answer_retention",
          "local_step": 1,
          "step": 5,
          "loss": 2.809854745864868,
          "trace_loss_active": false,
          "lr_scale": 1.0,
          "executor_loss": 2.809854745864868,
          "standard_len2_direct_accuracy": NaN,
          "standard_len2_direct_target_mass": NaN,
          "standard_len2_executor_accuracy": 0.0,
          "standard_len2_executor_target_mass": 0.061816165844599404,
          "standard_len2_init_accuracy": 0.08333333333333333,
          "standard_len2_op_accuracy": 0.4583333333333333,
          "standard_len2_arg_accuracy": 0.125,
          "standard_len2_program_exact": 0.0,
          "standard_len4_direct_accuracy": NaN,
          "standard_len4_direct_target_mass": NaN,
          "standard_len4_executor_accuracy": 0.0,
          "standard_len4_executor_target_mass": 0.05667603015899658,
          "standard_len4_init_accuracy": 0.0,
          "standard_len4_op_accuracy": 0.2708333333333333,
          "standard_len4_arg_accuracy": 0.08333333333333333,
          "standard_len4_program_exact": 0.0,
          "paraphrase_len2_direct_accuracy": NaN,
          "paraphrase_len2_direct_target_mass": NaN,
          "paraphrase_len2_executor_accuracy": 0.16666666666666666,
          "paraphrase_len2_executor_target_mass": 0.05950211485226949,
          "paraphrase_len2_init_accuracy": 0.0,
          "paraphrase_len2_op_accuracy": 0.3333333333333333,
          "paraphrase_len2_arg_accuracy": 0.125,
          "paraphrase_len2_program_exact": 0.0,
          "paraphrase_len4_direct_accuracy": NaN,
          "paraphrase_len4_direct_target_mass": NaN,
          "paraphrase_len4_executor_accuracy": 0.16666666666666666,
          "paraphrase_len4_executor_target_mass": 0.059757813811302185,
          "paraphrase_len4_init_accuracy": 0.0,
          "paraphrase_len4_op_accuracy": 0.4375,
          "paraphrase_len4_arg_accuracy": 0.020833333333333332,
          "paraphrase_len4_program_exact": 0.0
        },
        {
          "stage": "answer_retention",
          "local_step": 2,
          "step": 6,
          "loss": 2.864173173904419,
          "trace_loss_active": false,
          "lr_scale": 1.0,
          "executor_loss": 2.864173173904419,
          "standard_len2_direct_accuracy": NaN,
          "standard_len2_direct_target_mass": NaN,
          "standard_len2_executor_accuracy": 0.0,
          "standard_len2_executor_target_mass": 0.061729878187179565,
          "standard_len2_init_accuracy": 0.25,
          "standard_len2_op_accuracy": 0.4583333333333333,
          "standard_len2_arg_accuracy": 0.125,
          "standard_len2_program_exact": 0.0,
          "standard_len4_direct_accuracy": NaN,
          "standard_len4_direct_target_mass": NaN,
          "standard_len4_executor_accuracy": 0.08333333333333333,
          "standard_len4_executor_target_mass": 0.05668912331263224,
          "standard_len4_init_accuracy": 0.08333333333333333,
          "standard_len4_op_accuracy": 0.2708333333333333,
          "standard_len4_arg_accuracy": 0.08333333333333333,
          "standard_len4_program_exact": 0.0,
          "paraphrase_len2_direct_accuracy": NaN,
          "paraphrase_len2_direct_target_mass": NaN,
          "paraphrase_len2_executor_accuracy": 0.16666666666666666,
          "paraphrase_len2_executor_target_mass": 0.05946773290634155,
          "paraphrase_len2_init_accuracy": 0.0,
          "paraphrase_len2_op_accuracy": 0.3333333333333333,
          "paraphrase_len2_arg_accuracy": 0.08333333333333333,
          "paraphrase_len2_program_exact": 0.0,
          "paraphrase_len4_direct_accuracy": NaN,
          "paraphrase_len4_direct_target_mass": NaN,
          "paraphrase_len4_executor_accuracy": 0.0,
          "paraphrase_len4_executor_target_mass": 0.05976716677347819,
          "paraphrase_len4_init_accuracy": 0.0,
          "paraphrase_len4_op_accuracy": 0.4375,
          "paraphrase_len4_arg_accuracy": 0.020833333333333332,
          "paraphrase_len4_program_exact": 0.0
        },
        {
          "stage": "answer_retention",
          "local_step": 4,
          "step": 8,
          "loss": 2.865469217300415,
          "trace_loss_active": false,
          "lr_scale": 1.0,
          "executor_loss": 2.865469217300415,
          "standard_len2_direct_accuracy": NaN,
          "standard_len2_direct_target_mass": NaN,
          "standard_len2_executor_accuracy": 0.0,
          "standard_len2_executor_target_mass": 0.061691537499427795,
          "standard_len2_init_accuracy": 0.16666666666666666,
          "standard_len2_op_accuracy": 0.4583333333333333,
          "standard_len2_arg_accuracy": 0.125,
          "standard_len2_program_exact": 0.0,
          "standard_len4_direct_accuracy": NaN,
          "standard_len4_direct_target_mass": NaN,
          "standard_len4_executor_accuracy": 0.08333333333333333,
          "standard_len4_executor_target_mass": 0.056605140368143715,
          "standard_len4_init_accuracy": 0.08333333333333333,
          "standard_len4_op_accuracy": 0.2708333333333333,
          "standard_len4_arg_accuracy": 0.08333333333333333,
          "standard_len4_program_exact": 0.0,
          "paraphrase_len2_direct_accuracy": NaN,
          "paraphrase_len2_direct_target_mass": NaN,
          "paraphrase_len2_executor_accuracy": 0.16666666666666666,
          "paraphrase_len2_executor_target_mass": 0.05944700539112091,
          "paraphrase_len2_init_accuracy": 0.0,
          "paraphrase_len2_op_accuracy": 0.3333333333333333,
          "paraphrase_len2_arg_accuracy": 0.125,
          "paraphrase_len2_program_exact": 0.0,
          "paraphrase_len4_direct_accuracy": NaN,
          "paraphrase_len4_direct_target_mass": NaN,
          "paraphrase_len4_executor_accuracy": 0.0,
          "paraphrase_len4_executor_target_mass": 0.059806366761525474,
          "paraphrase_len4_init_accuracy": 0.0,
          "paraphrase_len4_op_accuracy": 0.4375,
          "paraphrase_len4_arg_accuracy": 0.020833333333333332,
          "paraphrase_len4_program_exact": 0.0
        }
      ],
      "final_metrics": {
        "standard_len2": {
          "variant": "compiler_trace_then_answer",
          "n": 12.0,
          "direct_accuracy": NaN,
          "direct_target_mass": NaN,
          "executor_accuracy": 0.0,
          "executor_target_mass": 0.061691537499427795,
          "init_accuracy": 0.16666666666666666,
          "op_accuracy": 0.4583333333333333,
          "arg_accuracy": 0.125,
          "program_exact": 0.0
        },
        "standard_len4": {
          "variant": "compiler_trace_then_answer",
          "n": 12.0,
          "direct_accuracy": NaN,
          "direct_target_mass": NaN,
          "executor_accuracy": 0.08333333333333333,
          "executor_target_mass": 0.056605140368143715,
          "init_accuracy": 0.08333333333333333,
          "op_accuracy": 0.2708333333333333,
          "arg_accuracy": 0.08333333333333333,
          "program_exact": 0.0
        },
        "paraphrase_len2": {
          "variant": "compiler_trace_then_answer",
          "n": 12.0,
          "direct_accuracy": NaN,
          "direct_target_mass": NaN,
          "executor_accuracy": 0.16666666666666666,
          "executor_target_mass": 0.05944700539112091,
          "init_accuracy": 0.0,
          "op_accuracy": 0.3333333333333333,
          "arg_accuracy": 0.125,
          "program_exact": 0.0
        },
        "paraphrase_len4": {
          "variant": "compiler_trace_then_answer",
          "n": 12.0,
          "direct_accuracy": NaN,
          "direct_target_mass": NaN,
          "executor_accuracy": 0.0,
          "executor_target_mass": 0.059806366761525474,
          "init_accuracy": 0.0,
          "op_accuracy": 0.4375,
          "arg_accuracy": 0.020833333333333332,
          "program_exact": 0.0
        }
      }
    }
  }
}