{
  "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": 12,
    "train_min_len": 1,
    "train_max_len": 3,
    "answer_train_min_len": 1,
    "answer_train_max_len": 6,
    "eval_lengths": "3,6,12",
    "train_template_mode": "standard",
    "answer_train_template_mode": "standard",
    "eval_template_modes": "standard,paraphrase",
    "train_size": 256,
    "answer_train_size": 256,
    "eval_size": 64,
    "feature_batch_size": 8,
    "train_batch_size": 64,
    "eval_batch_size": 64,
    "bootstrap_steps": 240,
    "answer_steps": 240,
    "eval_every": 120,
    "stage_eval_every": 120,
    "lr": 0.001,
    "answer_lr_scale": 0.1,
    "weight_decay": 0.01,
    "grad_clip": 1.0,
    "head_width": 384,
    "trace_loss_weight": 1.0,
    "executor_loss_weight": 1.0,
    "variants": "direct,compiler_trace,compiler_answer_only,compiler_trace_then_answer",
    "max_length": 512,
    "feature_seq_len": 192,
    "seed": 7,
    "output_dir": "experiments/qwen_span_free_compiler/runs/pilot_qwen35_span_free",
    "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": 256,
    "answer_train_size": 256,
    "eval_size": 64,
    "bootstrap_train_lengths": [
      1,
      3
    ],
    "answer_train_lengths": [
      1,
      6
    ],
    "eval_lengths": [
      3,
      6,
      12
    ],
    "train_template_mode": "standard",
    "answer_train_template_mode": "standard",
    "eval_template_modes": [
      "standard",
      "paraphrase"
    ],
    "feature_seq_len": 192
  },
  "variants": {
    "direct": {
      "variant": "direct",
      "train_seconds": 29.7902352809906,
      "checkpoints": [
        "large_artifacts/qwen_span_free_compiler/checkpoints/pilot_qwen35_span_free/direct/checkpoint_final.pt"
      ],
      "train_log": [
        {
          "stage": "answer_only",
          "local_step": 1,
          "step": 1,
          "loss": 4.562553405761719,
          "trace_loss_active": false,
          "lr_scale": 1.0,
          "direct_loss": 4.562553405761719,
          "standard_len3_direct_accuracy": 0.015625,
          "standard_len3_direct_target_mass": 0.011163948103785515,
          "standard_len3_executor_accuracy": NaN,
          "standard_len3_executor_target_mass": NaN,
          "standard_len3_init_accuracy": NaN,
          "standard_len3_op_accuracy": NaN,
          "standard_len3_arg_accuracy": NaN,
          "standard_len3_program_exact": NaN,
          "standard_len6_direct_accuracy": 0.015625,
          "standard_len6_direct_target_mass": 0.009834447875618935,
          "standard_len6_executor_accuracy": NaN,
          "standard_len6_executor_target_mass": NaN,
          "standard_len6_init_accuracy": NaN,
          "standard_len6_op_accuracy": NaN,
          "standard_len6_arg_accuracy": NaN,
          "standard_len6_program_exact": NaN,
          "standard_len12_direct_accuracy": 0.015625,
          "standard_len12_direct_target_mass": 0.011695163324475288,
          "standard_len12_executor_accuracy": NaN,
          "standard_len12_executor_target_mass": NaN,
          "standard_len12_init_accuracy": NaN,
          "standard_len12_op_accuracy": NaN,
          "standard_len12_arg_accuracy": NaN,
          "standard_len12_program_exact": NaN,
          "paraphrase_len3_direct_accuracy": 0.015625,
          "paraphrase_len3_direct_target_mass": 0.014907078817486763,
          "paraphrase_len3_executor_accuracy": NaN,
          "paraphrase_len3_executor_target_mass": NaN,
          "paraphrase_len3_init_accuracy": NaN,
          "paraphrase_len3_op_accuracy": NaN,
          "paraphrase_len3_arg_accuracy": NaN,
          "paraphrase_len3_program_exact": NaN,
          "paraphrase_len6_direct_accuracy": 0.0,
          "paraphrase_len6_direct_target_mass": 0.011500444263219833,
          "paraphrase_len6_executor_accuracy": NaN,
          "paraphrase_len6_executor_target_mass": NaN,
          "paraphrase_len6_init_accuracy": NaN,
          "paraphrase_len6_op_accuracy": NaN,
          "paraphrase_len6_arg_accuracy": NaN,
          "paraphrase_len6_program_exact": NaN,
          "paraphrase_len12_direct_accuracy": 0.0,
          "paraphrase_len12_direct_target_mass": 0.009197982028126717,
          "paraphrase_len12_executor_accuracy": NaN,
          "paraphrase_len12_executor_target_mass": NaN,
          "paraphrase_len12_init_accuracy": NaN,
          "paraphrase_len12_op_accuracy": NaN,
          "paraphrase_len12_arg_accuracy": NaN,
          "paraphrase_len12_program_exact": NaN
        },
        {
          "stage": "answer_only",
          "local_step": 120,
          "step": 120,
          "loss": 3.819924831390381,
          "trace_loss_active": false,
          "lr_scale": 1.0,
          "direct_loss": 3.819924831390381,
          "standard_len3_direct_accuracy": 0.0,
          "standard_len3_direct_target_mass": 0.010141398757696152,
          "standard_len3_executor_accuracy": NaN,
          "standard_len3_executor_target_mass": NaN,
          "standard_len3_init_accuracy": NaN,
          "standard_len3_op_accuracy": NaN,
          "standard_len3_arg_accuracy": NaN,
          "standard_len3_program_exact": NaN,
          "standard_len6_direct_accuracy": 0.0,
          "standard_len6_direct_target_mass": 0.009138772264122963,
          "standard_len6_executor_accuracy": NaN,
          "standard_len6_executor_target_mass": NaN,
          "standard_len6_init_accuracy": NaN,
          "standard_len6_op_accuracy": NaN,
          "standard_len6_arg_accuracy": NaN,
          "standard_len6_program_exact": NaN,
          "standard_len12_direct_accuracy": 0.0,
          "standard_len12_direct_target_mass": 0.009906514547765255,
          "standard_len12_executor_accuracy": NaN,
          "standard_len12_executor_target_mass": NaN,
          "standard_len12_init_accuracy": NaN,
          "standard_len12_op_accuracy": NaN,
          "standard_len12_arg_accuracy": NaN,
          "standard_len12_program_exact": NaN,
          "paraphrase_len3_direct_accuracy": 0.015625,
          "paraphrase_len3_direct_target_mass": 0.011593228206038475,
          "paraphrase_len3_executor_accuracy": NaN,
          "paraphrase_len3_executor_target_mass": NaN,
          "paraphrase_len3_init_accuracy": NaN,
          "paraphrase_len3_op_accuracy": NaN,
          "paraphrase_len3_arg_accuracy": NaN,
          "paraphrase_len3_program_exact": NaN,
          "paraphrase_len6_direct_accuracy": 0.0,
          "paraphrase_len6_direct_target_mass": 0.00876545999199152,
          "paraphrase_len6_executor_accuracy": NaN,
          "paraphrase_len6_executor_target_mass": NaN,
          "paraphrase_len6_init_accuracy": NaN,
          "paraphrase_len6_op_accuracy": NaN,
          "paraphrase_len6_arg_accuracy": NaN,
          "paraphrase_len6_program_exact": NaN,
          "paraphrase_len12_direct_accuracy": 0.015625,
          "paraphrase_len12_direct_target_mass": 0.011628970503807068,
          "paraphrase_len12_executor_accuracy": NaN,
          "paraphrase_len12_executor_target_mass": NaN,
          "paraphrase_len12_init_accuracy": NaN,
          "paraphrase_len12_op_accuracy": NaN,
          "paraphrase_len12_arg_accuracy": NaN,
          "paraphrase_len12_program_exact": NaN
        },
        {
          "stage": "answer_only",
          "local_step": 240,
          "step": 240,
          "loss": 3.554509162902832,
          "trace_loss_active": false,
          "lr_scale": 1.0,
          "direct_loss": 3.554509162902832,
          "standard_len3_direct_accuracy": 0.0,
          "standard_len3_direct_target_mass": 0.010846605524420738,
          "standard_len3_executor_accuracy": NaN,
          "standard_len3_executor_target_mass": NaN,
          "standard_len3_init_accuracy": NaN,
          "standard_len3_op_accuracy": NaN,
          "standard_len3_arg_accuracy": NaN,
          "standard_len3_program_exact": NaN,
          "standard_len6_direct_accuracy": 0.0,
          "standard_len6_direct_target_mass": 0.007259302772581577,
          "standard_len6_executor_accuracy": NaN,
          "standard_len6_executor_target_mass": NaN,
          "standard_len6_init_accuracy": NaN,
          "standard_len6_op_accuracy": NaN,
          "standard_len6_arg_accuracy": NaN,
          "standard_len6_program_exact": NaN,
          "standard_len12_direct_accuracy": 0.0,
          "standard_len12_direct_target_mass": 0.010653749108314514,
          "standard_len12_executor_accuracy": NaN,
          "standard_len12_executor_target_mass": NaN,
          "standard_len12_init_accuracy": NaN,
          "standard_len12_op_accuracy": NaN,
          "standard_len12_arg_accuracy": NaN,
          "standard_len12_program_exact": NaN,
          "paraphrase_len3_direct_accuracy": 0.015625,
          "paraphrase_len3_direct_target_mass": 0.011072760447859764,
          "paraphrase_len3_executor_accuracy": NaN,
          "paraphrase_len3_executor_target_mass": NaN,
          "paraphrase_len3_init_accuracy": NaN,
          "paraphrase_len3_op_accuracy": NaN,
          "paraphrase_len3_arg_accuracy": NaN,
          "paraphrase_len3_program_exact": NaN,
          "paraphrase_len6_direct_accuracy": 0.0,
          "paraphrase_len6_direct_target_mass": 0.008227115496993065,
          "paraphrase_len6_executor_accuracy": NaN,
          "paraphrase_len6_executor_target_mass": NaN,
          "paraphrase_len6_init_accuracy": NaN,
          "paraphrase_len6_op_accuracy": NaN,
          "paraphrase_len6_arg_accuracy": NaN,
          "paraphrase_len6_program_exact": NaN,
          "paraphrase_len12_direct_accuracy": 0.046875,
          "paraphrase_len12_direct_target_mass": 0.015281389467418194,
          "paraphrase_len12_executor_accuracy": NaN,
          "paraphrase_len12_executor_target_mass": NaN,
          "paraphrase_len12_init_accuracy": NaN,
          "paraphrase_len12_op_accuracy": NaN,
          "paraphrase_len12_arg_accuracy": NaN,
          "paraphrase_len12_program_exact": NaN
        },
        {
          "stage": "answer_only",
          "local_step": 360,
          "step": 360,
          "loss": 3.234121799468994,
          "trace_loss_active": false,
          "lr_scale": 1.0,
          "direct_loss": 3.234121799468994,
          "standard_len3_direct_accuracy": 0.0,
          "standard_len3_direct_target_mass": 0.010404914617538452,
          "standard_len3_executor_accuracy": NaN,
          "standard_len3_executor_target_mass": NaN,
          "standard_len3_init_accuracy": NaN,
          "standard_len3_op_accuracy": NaN,
          "standard_len3_arg_accuracy": NaN,
          "standard_len3_program_exact": NaN,
          "standard_len6_direct_accuracy": 0.0,
          "standard_len6_direct_target_mass": 0.00781078077852726,
          "standard_len6_executor_accuracy": NaN,
          "standard_len6_executor_target_mass": NaN,
          "standard_len6_init_accuracy": NaN,
          "standard_len6_op_accuracy": NaN,
          "standard_len6_arg_accuracy": NaN,
          "standard_len6_program_exact": NaN,
          "standard_len12_direct_accuracy": 0.046875,
          "standard_len12_direct_target_mass": 0.011398802511394024,
          "standard_len12_executor_accuracy": NaN,
          "standard_len12_executor_target_mass": NaN,
          "standard_len12_init_accuracy": NaN,
          "standard_len12_op_accuracy": NaN,
          "standard_len12_arg_accuracy": NaN,
          "standard_len12_program_exact": NaN,
          "paraphrase_len3_direct_accuracy": 0.015625,
          "paraphrase_len3_direct_target_mass": 0.014709116891026497,
          "paraphrase_len3_executor_accuracy": NaN,
          "paraphrase_len3_executor_target_mass": NaN,
          "paraphrase_len3_init_accuracy": NaN,
          "paraphrase_len3_op_accuracy": NaN,
          "paraphrase_len3_arg_accuracy": NaN,
          "paraphrase_len3_program_exact": NaN,
          "paraphrase_len6_direct_accuracy": 0.0,
          "paraphrase_len6_direct_target_mass": 0.006970865651965141,
          "paraphrase_len6_executor_accuracy": NaN,
          "paraphrase_len6_executor_target_mass": NaN,
          "paraphrase_len6_init_accuracy": NaN,
          "paraphrase_len6_op_accuracy": NaN,
          "paraphrase_len6_arg_accuracy": NaN,
          "paraphrase_len6_program_exact": NaN,
          "paraphrase_len12_direct_accuracy": 0.03125,
          "paraphrase_len12_direct_target_mass": 0.018483759835362434,
          "paraphrase_len12_executor_accuracy": NaN,
          "paraphrase_len12_executor_target_mass": NaN,
          "paraphrase_len12_init_accuracy": NaN,
          "paraphrase_len12_op_accuracy": NaN,
          "paraphrase_len12_arg_accuracy": NaN,
          "paraphrase_len12_program_exact": NaN
        },
        {
          "stage": "answer_only",
          "local_step": 480,
          "step": 480,
          "loss": 2.2494845390319824,
          "trace_loss_active": false,
          "lr_scale": 1.0,
          "direct_loss": 2.2494845390319824,
          "standard_len3_direct_accuracy": 0.015625,
          "standard_len3_direct_target_mass": 0.010317979380488396,
          "standard_len3_executor_accuracy": NaN,
          "standard_len3_executor_target_mass": NaN,
          "standard_len3_init_accuracy": NaN,
          "standard_len3_op_accuracy": NaN,
          "standard_len3_arg_accuracy": NaN,
          "standard_len3_program_exact": NaN,
          "standard_len6_direct_accuracy": 0.0,
          "standard_len6_direct_target_mass": 0.0068910615518689156,
          "standard_len6_executor_accuracy": NaN,
          "standard_len6_executor_target_mass": NaN,
          "standard_len6_init_accuracy": NaN,
          "standard_len6_op_accuracy": NaN,
          "standard_len6_arg_accuracy": NaN,
          "standard_len6_program_exact": NaN,
          "standard_len12_direct_accuracy": 0.015625,
          "standard_len12_direct_target_mass": 0.009835564531385899,
          "standard_len12_executor_accuracy": NaN,
          "standard_len12_executor_target_mass": NaN,
          "standard_len12_init_accuracy": NaN,
          "standard_len12_op_accuracy": NaN,
          "standard_len12_arg_accuracy": NaN,
          "standard_len12_program_exact": NaN,
          "paraphrase_len3_direct_accuracy": 0.015625,
          "paraphrase_len3_direct_target_mass": 0.01399741880595684,
          "paraphrase_len3_executor_accuracy": NaN,
          "paraphrase_len3_executor_target_mass": NaN,
          "paraphrase_len3_init_accuracy": NaN,
          "paraphrase_len3_op_accuracy": NaN,
          "paraphrase_len3_arg_accuracy": NaN,
          "paraphrase_len3_program_exact": NaN,
          "paraphrase_len6_direct_accuracy": 0.015625,
          "paraphrase_len6_direct_target_mass": 0.005514564923942089,
          "paraphrase_len6_executor_accuracy": NaN,
          "paraphrase_len6_executor_target_mass": NaN,
          "paraphrase_len6_init_accuracy": NaN,
          "paraphrase_len6_op_accuracy": NaN,
          "paraphrase_len6_arg_accuracy": NaN,
          "paraphrase_len6_program_exact": NaN,
          "paraphrase_len12_direct_accuracy": 0.03125,
          "paraphrase_len12_direct_target_mass": 0.023730961605906487,
          "paraphrase_len12_executor_accuracy": NaN,
          "paraphrase_len12_executor_target_mass": NaN,
          "paraphrase_len12_init_accuracy": NaN,
          "paraphrase_len12_op_accuracy": NaN,
          "paraphrase_len12_arg_accuracy": NaN,
          "paraphrase_len12_program_exact": NaN
        }
      ],
      "final_metrics": {
        "standard_len3": {
          "variant": "direct",
          "n": 64.0,
          "direct_accuracy": 0.015625,
          "direct_target_mass": 0.010317979380488396,
          "executor_accuracy": NaN,
          "executor_target_mass": NaN,
          "init_accuracy": NaN,
          "op_accuracy": NaN,
          "arg_accuracy": NaN,
          "program_exact": NaN
        },
        "standard_len6": {
          "variant": "direct",
          "n": 64.0,
          "direct_accuracy": 0.0,
          "direct_target_mass": 0.0068910615518689156,
          "executor_accuracy": NaN,
          "executor_target_mass": NaN,
          "init_accuracy": NaN,
          "op_accuracy": NaN,
          "arg_accuracy": NaN,
          "program_exact": NaN
        },
        "standard_len12": {
          "variant": "direct",
          "n": 64.0,
          "direct_accuracy": 0.015625,
          "direct_target_mass": 0.009835564531385899,
          "executor_accuracy": NaN,
          "executor_target_mass": NaN,
          "init_accuracy": NaN,
          "op_accuracy": NaN,
          "arg_accuracy": NaN,
          "program_exact": NaN
        },
        "paraphrase_len3": {
          "variant": "direct",
          "n": 64.0,
          "direct_accuracy": 0.015625,
          "direct_target_mass": 0.01399741880595684,
          "executor_accuracy": NaN,
          "executor_target_mass": NaN,
          "init_accuracy": NaN,
          "op_accuracy": NaN,
          "arg_accuracy": NaN,
          "program_exact": NaN
        },
        "paraphrase_len6": {
          "variant": "direct",
          "n": 64.0,
          "direct_accuracy": 0.015625,
          "direct_target_mass": 0.005514564923942089,
          "executor_accuracy": NaN,
          "executor_target_mass": NaN,
          "init_accuracy": NaN,
          "op_accuracy": NaN,
          "arg_accuracy": NaN,
          "program_exact": NaN
        },
        "paraphrase_len12": {
          "variant": "direct",
          "n": 64.0,
          "direct_accuracy": 0.03125,
          "direct_target_mass": 0.023730961605906487,
          "executor_accuracy": NaN,
          "executor_target_mass": NaN,
          "init_accuracy": NaN,
          "op_accuracy": NaN,
          "arg_accuracy": NaN,
          "program_exact": NaN
        }
      }
    },
    "compiler_trace": {
      "variant": "compiler_trace",
      "train_seconds": 38.78261995315552,
      "checkpoints": [
        "large_artifacts/qwen_span_free_compiler/checkpoints/pilot_qwen35_span_free/compiler_trace/checkpoint_final.pt"
      ],
      "train_log": [
        {
          "stage": "trace_bootstrap",
          "local_step": 1,
          "step": 1,
          "loss": 14.879148483276367,
          "trace_loss_active": true,
          "lr_scale": 1.0,
          "executor_loss": 4.562019348144531,
          "init_loss": 4.562651634216309,
          "op_loss": 1.1045829057693481,
          "arg_loss": 4.649894714355469,
          "standard_len3_direct_accuracy": NaN,
          "standard_len3_direct_target_mass": NaN,
          "standard_len3_executor_accuracy": 0.0,
          "standard_len3_executor_target_mass": 0.010277500376105309,
          "standard_len3_init_accuracy": 0.0,
          "standard_len3_op_accuracy": 0.2916666666666667,
          "standard_len3_arg_accuracy": 0.03125,
          "standard_len3_program_exact": 0.0,
          "standard_len6_direct_accuracy": NaN,
          "standard_len6_direct_target_mass": NaN,
          "standard_len6_executor_accuracy": 0.015625,
          "standard_len6_executor_target_mass": 0.010327626951038837,
          "standard_len6_init_accuracy": 0.0,
          "standard_len6_op_accuracy": 0.3515625,
          "standard_len6_arg_accuracy": 0.044270833333333336,
          "standard_len6_program_exact": 0.0,
          "standard_len12_direct_accuracy": NaN,
          "standard_len12_direct_target_mass": NaN,
          "standard_len12_executor_accuracy": 0.015625,
          "standard_len12_executor_target_mass": 0.010272540152072906,
          "standard_len12_init_accuracy": 0.03125,
          "standard_len12_op_accuracy": 0.3177083333333333,
          "standard_len12_arg_accuracy": 0.044270833333333336,
          "standard_len12_program_exact": 0.0,
          "paraphrase_len3_direct_accuracy": NaN,
          "paraphrase_len3_direct_target_mass": NaN,
          "paraphrase_len3_executor_accuracy": 0.015625,
          "paraphrase_len3_executor_target_mass": 0.010327920317649841,
          "paraphrase_len3_init_accuracy": 0.015625,
          "paraphrase_len3_op_accuracy": 0.3489583333333333,
          "paraphrase_len3_arg_accuracy": 0.06770833333333333,
          "paraphrase_len3_program_exact": 0.0,
          "paraphrase_len6_direct_accuracy": NaN,
          "paraphrase_len6_direct_target_mass": NaN,
          "paraphrase_len6_executor_accuracy": 0.015625,
          "paraphrase_len6_executor_target_mass": 0.010323813185095787,
          "paraphrase_len6_init_accuracy": 0.0,
          "paraphrase_len6_op_accuracy": 0.3177083333333333,
          "paraphrase_len6_arg_accuracy": 0.046875,
          "paraphrase_len6_program_exact": 0.0,
          "paraphrase_len12_direct_accuracy": NaN,
          "paraphrase_len12_direct_target_mass": NaN,
          "paraphrase_len12_executor_accuracy": 0.0,
          "paraphrase_len12_executor_target_mass": 0.010275041684508324,
          "paraphrase_len12_init_accuracy": 0.015625,
          "paraphrase_len12_op_accuracy": 0.3111979166666667,
          "paraphrase_len12_arg_accuracy": 0.0390625,
          "paraphrase_len12_program_exact": 0.0
        },
        {
          "stage": "trace_bootstrap",
          "local_step": 120,
          "step": 120,
          "loss": 11.075662612915039,
          "trace_loss_active": true,
          "lr_scale": 1.0,
          "executor_loss": 4.542809963226318,
          "init_loss": 2.570601463317871,
          "op_loss": 0.6060271263122559,
          "arg_loss": 3.3562238216400146,
          "standard_len3_direct_accuracy": NaN,
          "standard_len3_direct_target_mass": NaN,
          "standard_len3_executor_accuracy": 0.0,
          "standard_len3_executor_target_mass": 0.010356630198657513,
          "standard_len3_init_accuracy": 0.015625,
          "standard_len3_op_accuracy": 0.7916666666666666,
          "standard_len3_arg_accuracy": 0.026041666666666668,
          "standard_len3_program_exact": 0.0,
          "standard_len6_direct_accuracy": NaN,
          "standard_len6_direct_target_mass": NaN,
          "standard_len6_executor_accuracy": 0.015625,
          "standard_len6_executor_target_mass": 0.010295510292053223,
          "standard_len6_init_accuracy": 0.0,
          "standard_len6_op_accuracy": 0.4192708333333333,
          "standard_len6_arg_accuracy": 0.049479166666666664,
          "standard_len6_program_exact": 0.0,
          "standard_len12_direct_accuracy": NaN,
          "standard_len12_direct_target_mass": NaN,
          "standard_len12_executor_accuracy": 0.015625,
          "standard_len12_executor_target_mass": 0.010307235643267632,
          "standard_len12_init_accuracy": 0.0,
          "standard_len12_op_accuracy": 0.41796875,
          "standard_len12_arg_accuracy": 0.049479166666666664,
          "standard_len12_program_exact": 0.0,
          "paraphrase_len3_direct_accuracy": NaN,
          "paraphrase_len3_direct_target_mass": NaN,
          "paraphrase_len3_executor_accuracy": 0.0,
          "paraphrase_len3_executor_target_mass": 0.010268054902553558,
          "paraphrase_len3_init_accuracy": 0.03125,
          "paraphrase_len3_op_accuracy": 0.4583333333333333,
          "paraphrase_len3_arg_accuracy": 0.041666666666666664,
          "paraphrase_len3_program_exact": 0.0,
          "paraphrase_len6_direct_accuracy": NaN,
          "paraphrase_len6_direct_target_mass": NaN,
          "paraphrase_len6_executor_accuracy": 0.015625,
          "paraphrase_len6_executor_target_mass": 0.010291149839758873,
          "paraphrase_len6_init_accuracy": 0.046875,
          "paraphrase_len6_op_accuracy": 0.328125,
          "paraphrase_len6_arg_accuracy": 0.052083333333333336,
          "paraphrase_len6_program_exact": 0.0,
          "paraphrase_len12_direct_accuracy": NaN,
          "paraphrase_len12_direct_target_mass": NaN,
          "paraphrase_len12_executor_accuracy": 0.03125,
          "paraphrase_len12_executor_target_mass": 0.010261422023177147,
          "paraphrase_len12_init_accuracy": 0.0,
          "paraphrase_len12_op_accuracy": 0.3268229166666667,
          "paraphrase_len12_arg_accuracy": 0.059895833333333336,
          "paraphrase_len12_program_exact": 0.0
        },
        {
          "stage": "trace_bootstrap",
          "local_step": 240,
          "step": 240,
          "loss": 9.541241645812988,
          "trace_loss_active": true,
          "lr_scale": 1.0,
          "executor_loss": 4.455385208129883,
          "init_loss": 1.045335054397583,
          "op_loss": 0.6023046970367432,
          "arg_loss": 3.4382164478302,
          "standard_len3_direct_accuracy": NaN,
          "standard_len3_direct_target_mass": NaN,
          "standard_len3_executor_accuracy": 0.0625,
          "standard_len3_executor_target_mass": 0.010183983482420444,
          "standard_len3_init_accuracy": 0.03125,
          "standard_len3_op_accuracy": 0.7864583333333334,
          "standard_len3_arg_accuracy": 0.03125,
          "standard_len3_program_exact": 0.0,
          "standard_len6_direct_accuracy": NaN,
          "standard_len6_direct_target_mass": NaN,
          "standard_len6_executor_accuracy": 0.015625,
          "standard_len6_executor_target_mass": 0.010287151671946049,
          "standard_len6_init_accuracy": 0.015625,
          "standard_len6_op_accuracy": 0.4557291666666667,
          "standard_len6_arg_accuracy": 0.041666666666666664,
          "standard_len6_program_exact": 0.0,
          "standard_len12_direct_accuracy": NaN,
          "standard_len12_direct_target_mass": NaN,
          "standard_len12_executor_accuracy": 0.046875,
          "standard_len12_executor_target_mass": 0.01030139159411192,
          "standard_len12_init_accuracy": 0.015625,
          "standard_len12_op_accuracy": 0.4635416666666667,
          "standard_len12_arg_accuracy": 0.046875,
          "standard_len12_program_exact": 0.0,
          "paraphrase_len3_direct_accuracy": NaN,
          "paraphrase_len3_direct_target_mass": NaN,
          "paraphrase_len3_executor_accuracy": 0.0,
          "paraphrase_len3_executor_target_mass": 0.010511185973882675,
          "paraphrase_len3_init_accuracy": 0.0,
          "paraphrase_len3_op_accuracy": 0.4322916666666667,
          "paraphrase_len3_arg_accuracy": 0.052083333333333336,
          "paraphrase_len3_program_exact": 0.0,
          "paraphrase_len6_direct_accuracy": NaN,
          "paraphrase_len6_direct_target_mass": NaN,
          "paraphrase_len6_executor_accuracy": 0.0,
          "paraphrase_len6_executor_target_mass": 0.010266849771142006,
          "paraphrase_len6_init_accuracy": 0.03125,
          "paraphrase_len6_op_accuracy": 0.3463541666666667,
          "paraphrase_len6_arg_accuracy": 0.0390625,
          "paraphrase_len6_program_exact": 0.0,
          "paraphrase_len12_direct_accuracy": NaN,
          "paraphrase_len12_direct_target_mass": NaN,
          "paraphrase_len12_executor_accuracy": 0.015625,
          "paraphrase_len12_executor_target_mass": 0.010248461738228798,
          "paraphrase_len12_init_accuracy": 0.0,
          "paraphrase_len12_op_accuracy": 0.3294270833333333,
          "paraphrase_len12_arg_accuracy": 0.0390625,
          "paraphrase_len12_program_exact": 0.0
        },
        {
          "stage": "trace_continuation",
          "local_step": 1,
          "step": 241,
          "loss": 20.824600219726562,
          "trace_loss_active": true,
          "lr_scale": 1.0,
          "executor_loss": 4.606600761413574,
          "init_loss": 11.238036155700684,
          "op_loss": 1.252516508102417,
          "arg_loss": 3.7274465560913086,
          "standard_len3_direct_accuracy": NaN,
          "standard_len3_direct_target_mass": NaN,
          "standard_len3_executor_accuracy": 0.015625,
          "standard_len3_executor_target_mass": 0.010104052722454071,
          "standard_len3_init_accuracy": 0.03125,
          "standard_len3_op_accuracy": 0.7864583333333334,
          "standard_len3_arg_accuracy": 0.03125,
          "standard_len3_program_exact": 0.0,
          "standard_len6_direct_accuracy": NaN,
          "standard_len6_direct_target_mass": NaN,
          "standard_len6_executor_accuracy": 0.03125,
          "standard_len6_executor_target_mass": 0.010302508249878883,
          "standard_len6_init_accuracy": 0.03125,
          "standard_len6_op_accuracy": 0.4713541666666667,
          "standard_len6_arg_accuracy": 0.046875,
          "standard_len6_program_exact": 0.0,
          "standard_len12_direct_accuracy": NaN,
          "standard_len12_direct_target_mass": NaN,
          "standard_len12_executor_accuracy": 0.015625,
          "standard_len12_executor_target_mass": 0.010287552140653133,
          "standard_len12_init_accuracy": 0.03125,
          "standard_len12_op_accuracy": 0.4388020833333333,
          "standard_len12_arg_accuracy": 0.052083333333333336,
          "standard_len12_program_exact": 0.0,
          "paraphrase_len3_direct_accuracy": NaN,
          "paraphrase_len3_direct_target_mass": NaN,
          "paraphrase_len3_executor_accuracy": 0.0,
          "paraphrase_len3_executor_target_mass": 0.010627612471580505,
          "paraphrase_len3_init_accuracy": 0.0,
          "paraphrase_len3_op_accuracy": 0.4322916666666667,
          "paraphrase_len3_arg_accuracy": 0.052083333333333336,
          "paraphrase_len3_program_exact": 0.0,
          "paraphrase_len6_direct_accuracy": NaN,
          "paraphrase_len6_direct_target_mass": NaN,
          "paraphrase_len6_executor_accuracy": 0.0,
          "paraphrase_len6_executor_target_mass": 0.010305508971214294,
          "paraphrase_len6_init_accuracy": 0.03125,
          "paraphrase_len6_op_accuracy": 0.3671875,
          "paraphrase_len6_arg_accuracy": 0.0546875,
          "paraphrase_len6_program_exact": 0.0,
          "paraphrase_len12_direct_accuracy": NaN,
          "paraphrase_len12_direct_target_mass": NaN,
          "paraphrase_len12_executor_accuracy": 0.015625,
          "paraphrase_len12_executor_target_mass": 0.01024479791522026,
          "paraphrase_len12_init_accuracy": 0.0,
          "paraphrase_len12_op_accuracy": 0.34375,
          "paraphrase_len12_arg_accuracy": 0.04296875,
          "paraphrase_len12_program_exact": 0.0
        },
        {
          "stage": "trace_continuation",
          "local_step": 120,
          "step": 360,
          "loss": 9.773493766784668,
          "trace_loss_active": true,
          "lr_scale": 1.0,
          "executor_loss": 4.526269912719727,
          "init_loss": 1.468888521194458,
          "op_loss": 0.4730089008808136,
          "arg_loss": 3.3053267002105713,
          "standard_len3_direct_accuracy": NaN,
          "standard_len3_direct_target_mass": NaN,
          "standard_len3_executor_accuracy": 0.03125,
          "standard_len3_executor_target_mass": 0.010282192379236221,
          "standard_len3_init_accuracy": 0.015625,
          "standard_len3_op_accuracy": 0.7552083333333334,
          "standard_len3_arg_accuracy": 0.036458333333333336,
          "standard_len3_program_exact": 0.0,
          "standard_len6_direct_accuracy": NaN,
          "standard_len6_direct_target_mass": NaN,
          "standard_len6_executor_accuracy": 0.015625,
          "standard_len6_executor_target_mass": 0.010342942550778389,
          "standard_len6_init_accuracy": 0.03125,
          "standard_len6_op_accuracy": 0.7057291666666666,
          "standard_len6_arg_accuracy": 0.06510416666666667,
          "standard_len6_program_exact": 0.0,
          "standard_len12_direct_accuracy": NaN,
          "standard_len12_direct_target_mass": NaN,
          "standard_len12_executor_accuracy": 0.015625,
          "standard_len12_executor_target_mass": 0.010297699831426144,
          "standard_len12_init_accuracy": 0.015625,
          "standard_len12_op_accuracy": 0.48828125,
          "standard_len12_arg_accuracy": 0.049479166666666664,
          "standard_len12_program_exact": 0.0,
          "paraphrase_len3_direct_accuracy": NaN,
          "paraphrase_len3_direct_target_mass": NaN,
          "paraphrase_len3_executor_accuracy": 0.0,
          "paraphrase_len3_executor_target_mass": 0.010245861485600471,
          "paraphrase_len3_init_accuracy": 0.015625,
          "paraphrase_len3_op_accuracy": 0.4479166666666667,
          "paraphrase_len3_arg_accuracy": 0.03125,
          "paraphrase_len3_program_exact": 0.0,
          "paraphrase_len6_direct_accuracy": NaN,
          "paraphrase_len6_direct_target_mass": NaN,
          "paraphrase_len6_executor_accuracy": 0.0,
          "paraphrase_len6_executor_target_mass": 0.010308504104614258,
          "paraphrase_len6_init_accuracy": 0.015625,
          "paraphrase_len6_op_accuracy": 0.3802083333333333,
          "paraphrase_len6_arg_accuracy": 0.041666666666666664,
          "paraphrase_len6_program_exact": 0.0,
          "paraphrase_len12_direct_accuracy": NaN,
          "paraphrase_len12_direct_target_mass": NaN,
          "paraphrase_len12_executor_accuracy": 0.015625,
          "paraphrase_len12_executor_target_mass": 0.010260218754410744,
          "paraphrase_len12_init_accuracy": 0.03125,
          "paraphrase_len12_op_accuracy": 0.3658854166666667,
          "paraphrase_len12_arg_accuracy": 0.036458333333333336,
          "paraphrase_len12_program_exact": 0.0
        },
        {
          "stage": "trace_continuation",
          "local_step": 240,
          "step": 480,
          "loss": 8.774296760559082,
          "trace_loss_active": true,
          "lr_scale": 1.0,
          "executor_loss": 4.535475254058838,
          "init_loss": 0.6178188323974609,
          "op_loss": 0.40945711731910706,
          "arg_loss": 3.211545705795288,
          "standard_len3_direct_accuracy": NaN,
          "standard_len3_direct_target_mass": NaN,
          "standard_len3_executor_accuracy": 0.015625,
          "standard_len3_executor_target_mass": 0.010481145232915878,
          "standard_len3_init_accuracy": 0.0,
          "standard_len3_op_accuracy": 0.7552083333333334,
          "standard_len3_arg_accuracy": 0.07291666666666667,
          "standard_len3_program_exact": 0.0,
          "standard_len6_direct_accuracy": NaN,
          "standard_len6_direct_target_mass": NaN,
          "standard_len6_executor_accuracy": 0.03125,
          "standard_len6_executor_target_mass": 0.010300541296601295,
          "standard_len6_init_accuracy": 0.046875,
          "standard_len6_op_accuracy": 0.7213541666666666,
          "standard_len6_arg_accuracy": 0.0625,
          "standard_len6_program_exact": 0.0,
          "standard_len12_direct_accuracy": NaN,
          "standard_len12_direct_target_mass": NaN,
          "standard_len12_executor_accuracy": 0.046875,
          "standard_len12_executor_target_mass": 0.010305428877472878,
          "standard_len12_init_accuracy": 0.015625,
          "standard_len12_op_accuracy": 0.48828125,
          "standard_len12_arg_accuracy": 0.0390625,
          "standard_len12_program_exact": 0.0,
          "paraphrase_len3_direct_accuracy": NaN,
          "paraphrase_len3_direct_target_mass": NaN,
          "paraphrase_len3_executor_accuracy": 0.0,
          "paraphrase_len3_executor_target_mass": 0.010308768600225449,
          "paraphrase_len3_init_accuracy": 0.015625,
          "paraphrase_len3_op_accuracy": 0.4375,
          "paraphrase_len3_arg_accuracy": 0.0625,
          "paraphrase_len3_program_exact": 0.0,
          "paraphrase_len6_direct_accuracy": NaN,
          "paraphrase_len6_direct_target_mass": NaN,
          "paraphrase_len6_executor_accuracy": 0.0,
          "paraphrase_len6_executor_target_mass": 0.010304001159965992,
          "paraphrase_len6_init_accuracy": 0.015625,
          "paraphrase_len6_op_accuracy": 0.3645833333333333,
          "paraphrase_len6_arg_accuracy": 0.036458333333333336,
          "paraphrase_len6_program_exact": 0.0,
          "paraphrase_len12_direct_accuracy": NaN,
          "paraphrase_len12_direct_target_mass": NaN,
          "paraphrase_len12_executor_accuracy": 0.015625,
          "paraphrase_len12_executor_target_mass": 0.010240824893116951,
          "paraphrase_len12_init_accuracy": 0.0,
          "paraphrase_len12_op_accuracy": 0.3776041666666667,
          "paraphrase_len12_arg_accuracy": 0.037760416666666664,
          "paraphrase_len12_program_exact": 0.0
        }
      ],
      "final_metrics": {
        "standard_len3": {
          "variant": "compiler_trace",
          "n": 64.0,
          "direct_accuracy": NaN,
          "direct_target_mass": NaN,
          "executor_accuracy": 0.015625,
          "executor_target_mass": 0.010481145232915878,
          "init_accuracy": 0.0,
          "op_accuracy": 0.7552083333333334,
          "arg_accuracy": 0.07291666666666667,
          "program_exact": 0.0
        },
        "standard_len6": {
          "variant": "compiler_trace",
          "n": 64.0,
          "direct_accuracy": NaN,
          "direct_target_mass": NaN,
          "executor_accuracy": 0.03125,
          "executor_target_mass": 0.010300541296601295,
          "init_accuracy": 0.046875,
          "op_accuracy": 0.7213541666666666,
          "arg_accuracy": 0.0625,
          "program_exact": 0.0
        },
        "standard_len12": {
          "variant": "compiler_trace",
          "n": 64.0,
          "direct_accuracy": NaN,
          "direct_target_mass": NaN,
          "executor_accuracy": 0.046875,
          "executor_target_mass": 0.010305428877472878,
          "init_accuracy": 0.015625,
          "op_accuracy": 0.48828125,
          "arg_accuracy": 0.0390625,
          "program_exact": 0.0
        },
        "paraphrase_len3": {
          "variant": "compiler_trace",
          "n": 64.0,
          "direct_accuracy": NaN,
          "direct_target_mass": NaN,
          "executor_accuracy": 0.0,
          "executor_target_mass": 0.010308768600225449,
          "init_accuracy": 0.015625,
          "op_accuracy": 0.4375,
          "arg_accuracy": 0.0625,
          "program_exact": 0.0
        },
        "paraphrase_len6": {
          "variant": "compiler_trace",
          "n": 64.0,
          "direct_accuracy": NaN,
          "direct_target_mass": NaN,
          "executor_accuracy": 0.0,
          "executor_target_mass": 0.010304001159965992,
          "init_accuracy": 0.015625,
          "op_accuracy": 0.3645833333333333,
          "arg_accuracy": 0.036458333333333336,
          "program_exact": 0.0
        },
        "paraphrase_len12": {
          "variant": "compiler_trace",
          "n": 64.0,
          "direct_accuracy": NaN,
          "direct_target_mass": NaN,
          "executor_accuracy": 0.015625,
          "executor_target_mass": 0.010240824893116951,
          "init_accuracy": 0.0,
          "op_accuracy": 0.3776041666666667,
          "arg_accuracy": 0.037760416666666664,
          "program_exact": 0.0
        }
      }
    },
    "compiler_answer_only": {
      "variant": "compiler_answer_only",
      "train_seconds": 38.78789019584656,
      "checkpoints": [
        "large_artifacts/qwen_span_free_compiler/checkpoints/pilot_qwen35_span_free/compiler_answer_only/checkpoint_final.pt"
      ],
      "train_log": [
        {
          "stage": "answer_only",
          "local_step": 1,
          "step": 1,
          "loss": 4.574073314666748,
          "trace_loss_active": false,
          "lr_scale": 1.0,
          "executor_loss": 4.574073314666748,
          "standard_len3_direct_accuracy": NaN,
          "standard_len3_direct_target_mass": NaN,
          "standard_len3_executor_accuracy": 0.015625,
          "standard_len3_executor_target_mass": 0.01030229777097702,
          "standard_len3_init_accuracy": 0.0,
          "standard_len3_op_accuracy": 0.3125,
          "standard_len3_arg_accuracy": 0.036458333333333336,
          "standard_len3_program_exact": 0.0,
          "standard_len6_direct_accuracy": NaN,
          "standard_len6_direct_target_mass": NaN,
          "standard_len6_executor_accuracy": 0.015625,
          "standard_len6_executor_target_mass": 0.010311576537787914,
          "standard_len6_init_accuracy": 0.015625,
          "standard_len6_op_accuracy": 0.2994791666666667,
          "standard_len6_arg_accuracy": 0.059895833333333336,
          "standard_len6_program_exact": 0.0,
          "standard_len12_direct_accuracy": NaN,
          "standard_len12_direct_target_mass": NaN,
          "standard_len12_executor_accuracy": 0.015625,
          "standard_len12_executor_target_mass": 0.010306401178240776,
          "standard_len12_init_accuracy": 0.0,
          "standard_len12_op_accuracy": 0.3515625,
          "standard_len12_arg_accuracy": 0.033854166666666664,
          "standard_len12_program_exact": 0.0,
          "paraphrase_len3_direct_accuracy": NaN,
          "paraphrase_len3_direct_target_mass": NaN,
          "paraphrase_len3_executor_accuracy": 0.03125,
          "paraphrase_len3_executor_target_mass": 0.01031481847167015,
          "paraphrase_len3_init_accuracy": 0.0,
          "paraphrase_len3_op_accuracy": 0.3385416666666667,
          "paraphrase_len3_arg_accuracy": 0.052083333333333336,
          "paraphrase_len3_program_exact": 0.0,
          "paraphrase_len6_direct_accuracy": NaN,
          "paraphrase_len6_direct_target_mass": NaN,
          "paraphrase_len6_executor_accuracy": 0.0,
          "paraphrase_len6_executor_target_mass": 0.010311335325241089,
          "paraphrase_len6_init_accuracy": 0.0,
          "paraphrase_len6_op_accuracy": 0.3125,
          "paraphrase_len6_arg_accuracy": 0.0625,
          "paraphrase_len6_program_exact": 0.0,
          "paraphrase_len12_direct_accuracy": NaN,
          "paraphrase_len12_direct_target_mass": NaN,
          "paraphrase_len12_executor_accuracy": 0.0,
          "paraphrase_len12_executor_target_mass": 0.010306239128112793,
          "paraphrase_len12_init_accuracy": 0.015625,
          "paraphrase_len12_op_accuracy": 0.3255208333333333,
          "paraphrase_len12_arg_accuracy": 0.049479166666666664,
          "paraphrase_len12_program_exact": 0.0
        },
        {
          "stage": "answer_only",
          "local_step": 120,
          "step": 120,
          "loss": 4.305878639221191,
          "trace_loss_active": false,
          "lr_scale": 1.0,
          "executor_loss": 4.305878639221191,
          "standard_len3_direct_accuracy": NaN,
          "standard_len3_direct_target_mass": NaN,
          "standard_len3_executor_accuracy": 0.0,
          "standard_len3_executor_target_mass": 0.012398120015859604,
          "standard_len3_init_accuracy": 0.0,
          "standard_len3_op_accuracy": 0.2916666666666667,
          "standard_len3_arg_accuracy": 0.0,
          "standard_len3_program_exact": 0.0,
          "standard_len6_direct_accuracy": NaN,
          "standard_len6_direct_target_mass": NaN,
          "standard_len6_executor_accuracy": 0.015625,
          "standard_len6_executor_target_mass": 0.0105275334790349,
          "standard_len6_init_accuracy": 0.015625,
          "standard_len6_op_accuracy": 0.3515625,
          "standard_len6_arg_accuracy": 0.0026041666666666665,
          "standard_len6_program_exact": 0.0,
          "standard_len12_direct_accuracy": NaN,
          "standard_len12_direct_target_mass": NaN,
          "standard_len12_executor_accuracy": 0.0,
          "standard_len12_executor_target_mass": 0.010100426152348518,
          "standard_len12_init_accuracy": 0.015625,
          "standard_len12_op_accuracy": 0.3177083333333333,
          "standard_len12_arg_accuracy": 0.005208333333333333,
          "standard_len12_program_exact": 0.0,
          "paraphrase_len3_direct_accuracy": NaN,
          "paraphrase_len3_direct_target_mass": NaN,
          "paraphrase_len3_executor_accuracy": 0.015625,
          "paraphrase_len3_executor_target_mass": 0.010943718254566193,
          "paraphrase_len3_init_accuracy": 0.0,
          "paraphrase_len3_op_accuracy": 0.3489583333333333,
          "paraphrase_len3_arg_accuracy": 0.0,
          "paraphrase_len3_program_exact": 0.0,
          "paraphrase_len6_direct_accuracy": NaN,
          "paraphrase_len6_direct_target_mass": NaN,
          "paraphrase_len6_executor_accuracy": 0.0,
          "paraphrase_len6_executor_target_mass": 0.010682889260351658,
          "paraphrase_len6_init_accuracy": 0.03125,
          "paraphrase_len6_op_accuracy": 0.3177083333333333,
          "paraphrase_len6_arg_accuracy": 0.0078125,
          "paraphrase_len6_program_exact": 0.0,
          "paraphrase_len12_direct_accuracy": NaN,
          "paraphrase_len12_direct_target_mass": NaN,
          "paraphrase_len12_executor_accuracy": 0.015625,
          "paraphrase_len12_executor_target_mass": 0.010432576760649681,
          "paraphrase_len12_init_accuracy": 0.015625,
          "paraphrase_len12_op_accuracy": 0.3111979166666667,
          "paraphrase_len12_arg_accuracy": 0.005208333333333333,
          "paraphrase_len12_program_exact": 0.0
        },
        {
          "stage": "answer_only",
          "local_step": 240,
          "step": 240,
          "loss": 3.5869648456573486,
          "trace_loss_active": false,
          "lr_scale": 1.0,
          "executor_loss": 3.5869648456573486,
          "standard_len3_direct_accuracy": NaN,
          "standard_len3_direct_target_mass": NaN,
          "standard_len3_executor_accuracy": 0.015625,
          "standard_len3_executor_target_mass": 0.012636525556445122,
          "standard_len3_init_accuracy": 0.046875,
          "standard_len3_op_accuracy": 0.2916666666666667,
          "standard_len3_arg_accuracy": 0.0,
          "standard_len3_program_exact": 0.0,
          "standard_len6_direct_accuracy": NaN,
          "standard_len6_direct_target_mass": NaN,
          "standard_len6_executor_accuracy": 0.0,
          "standard_len6_executor_target_mass": 0.010589079931378365,
          "standard_len6_init_accuracy": 0.015625,
          "standard_len6_op_accuracy": 0.3515625,
          "standard_len6_arg_accuracy": 0.0,
          "standard_len6_program_exact": 0.0,
          "standard_len12_direct_accuracy": NaN,
          "standard_len12_direct_target_mass": NaN,
          "standard_len12_executor_accuracy": 0.0,
          "standard_len12_executor_target_mass": 0.010113993659615517,
          "standard_len12_init_accuracy": 0.0,
          "standard_len12_op_accuracy": 0.3177083333333333,
          "standard_len12_arg_accuracy": 0.005208333333333333,
          "standard_len12_program_exact": 0.0,
          "paraphrase_len3_direct_accuracy": NaN,
          "paraphrase_len3_direct_target_mass": NaN,
          "paraphrase_len3_executor_accuracy": 0.03125,
          "paraphrase_len3_executor_target_mass": 0.01036088913679123,
          "paraphrase_len3_init_accuracy": 0.03125,
          "paraphrase_len3_op_accuracy": 0.3489583333333333,
          "paraphrase_len3_arg_accuracy": 0.0,
          "paraphrase_len3_program_exact": 0.0,
          "paraphrase_len6_direct_accuracy": NaN,
          "paraphrase_len6_direct_target_mass": NaN,
          "paraphrase_len6_executor_accuracy": 0.015625,
          "paraphrase_len6_executor_target_mass": 0.010504872538149357,
          "paraphrase_len6_init_accuracy": 0.015625,
          "paraphrase_len6_op_accuracy": 0.3177083333333333,
          "paraphrase_len6_arg_accuracy": 0.005208333333333333,
          "paraphrase_len6_program_exact": 0.0,
          "paraphrase_len12_direct_accuracy": NaN,
          "paraphrase_len12_direct_target_mass": NaN,
          "paraphrase_len12_executor_accuracy": 0.015625,
          "paraphrase_len12_executor_target_mass": 0.010121665894985199,
          "paraphrase_len12_init_accuracy": 0.015625,
          "paraphrase_len12_op_accuracy": 0.3111979166666667,
          "paraphrase_len12_arg_accuracy": 0.005208333333333333,
          "paraphrase_len12_program_exact": 0.0
        },
        {
          "stage": "answer_only",
          "local_step": 360,
          "step": 360,
          "loss": 3.026268243789673,
          "trace_loss_active": false,
          "lr_scale": 1.0,
          "executor_loss": 3.026268243789673,
          "standard_len3_direct_accuracy": NaN,
          "standard_len3_direct_target_mass": NaN,
          "standard_len3_executor_accuracy": 0.0,
          "standard_len3_executor_target_mass": 0.013241693377494812,
          "standard_len3_init_accuracy": 0.0,
          "standard_len3_op_accuracy": 0.2916666666666667,
          "standard_len3_arg_accuracy": 0.0,
          "standard_len3_program_exact": 0.0,
          "standard_len6_direct_accuracy": NaN,
          "standard_len6_direct_target_mass": NaN,
          "standard_len6_executor_accuracy": 0.0,
          "standard_len6_executor_target_mass": 0.010018456727266312,
          "standard_len6_init_accuracy": 0.015625,
          "standard_len6_op_accuracy": 0.3515625,
          "standard_len6_arg_accuracy": 0.0,
          "standard_len6_program_exact": 0.0,
          "standard_len12_direct_accuracy": NaN,
          "standard_len12_direct_target_mass": NaN,
          "standard_len12_executor_accuracy": 0.0,
          "standard_len12_executor_target_mass": 0.010780847631394863,
          "standard_len12_init_accuracy": 0.03125,
          "standard_len12_op_accuracy": 0.3177083333333333,
          "standard_len12_arg_accuracy": 0.006510416666666667,
          "standard_len12_program_exact": 0.0,
          "paraphrase_len3_direct_accuracy": NaN,
          "paraphrase_len3_direct_target_mass": NaN,
          "paraphrase_len3_executor_accuracy": 0.0,
          "paraphrase_len3_executor_target_mass": 0.008443934842944145,
          "paraphrase_len3_init_accuracy": 0.015625,
          "paraphrase_len3_op_accuracy": 0.3489583333333333,
          "paraphrase_len3_arg_accuracy": 0.0,
          "paraphrase_len3_program_exact": 0.0,
          "paraphrase_len6_direct_accuracy": NaN,
          "paraphrase_len6_direct_target_mass": NaN,
          "paraphrase_len6_executor_accuracy": 0.0,
          "paraphrase_len6_executor_target_mass": 0.00986553356051445,
          "paraphrase_len6_init_accuracy": 0.046875,
          "paraphrase_len6_op_accuracy": 0.3177083333333333,
          "paraphrase_len6_arg_accuracy": 0.0026041666666666665,
          "paraphrase_len6_program_exact": 0.0,
          "paraphrase_len12_direct_accuracy": NaN,
          "paraphrase_len12_direct_target_mass": NaN,
          "paraphrase_len12_executor_accuracy": 0.0,
          "paraphrase_len12_executor_target_mass": 0.00966302677989006,
          "paraphrase_len12_init_accuracy": 0.03125,
          "paraphrase_len12_op_accuracy": 0.3111979166666667,
          "paraphrase_len12_arg_accuracy": 0.0026041666666666665,
          "paraphrase_len12_program_exact": 0.0
        },
        {
          "stage": "answer_only",
          "local_step": 480,
          "step": 480,
          "loss": 2.678358793258667,
          "trace_loss_active": false,
          "lr_scale": 1.0,
          "executor_loss": 2.678358793258667,
          "standard_len3_direct_accuracy": NaN,
          "standard_len3_direct_target_mass": NaN,
          "standard_len3_executor_accuracy": 0.03125,
          "standard_len3_executor_target_mass": 0.015396549366414547,
          "standard_len3_init_accuracy": 0.0,
          "standard_len3_op_accuracy": 0.2916666666666667,
          "standard_len3_arg_accuracy": 0.0,
          "standard_len3_program_exact": 0.0,
          "standard_len6_direct_accuracy": NaN,
          "standard_len6_direct_target_mass": NaN,
          "standard_len6_executor_accuracy": 0.0,
          "standard_len6_executor_target_mass": 0.010013941675424576,
          "standard_len6_init_accuracy": 0.03125,
          "standard_len6_op_accuracy": 0.3515625,
          "standard_len6_arg_accuracy": 0.0,
          "standard_len6_program_exact": 0.0,
          "standard_len12_direct_accuracy": NaN,
          "standard_len12_direct_target_mass": NaN,
          "standard_len12_executor_accuracy": 0.0,
          "standard_len12_executor_target_mass": 0.010086162015795708,
          "standard_len12_init_accuracy": 0.015625,
          "standard_len12_op_accuracy": 0.3177083333333333,
          "standard_len12_arg_accuracy": 0.006510416666666667,
          "standard_len12_program_exact": 0.0,
          "paraphrase_len3_direct_accuracy": NaN,
          "paraphrase_len3_direct_target_mass": NaN,
          "paraphrase_len3_executor_accuracy": 0.0,
          "paraphrase_len3_executor_target_mass": 0.006057328544557095,
          "paraphrase_len3_init_accuracy": 0.015625,
          "paraphrase_len3_op_accuracy": 0.3489583333333333,
          "paraphrase_len3_arg_accuracy": 0.0,
          "paraphrase_len3_program_exact": 0.0,
          "paraphrase_len6_direct_accuracy": NaN,
          "paraphrase_len6_direct_target_mass": NaN,
          "paraphrase_len6_executor_accuracy": 0.015625,
          "paraphrase_len6_executor_target_mass": 0.008675915189087391,
          "paraphrase_len6_init_accuracy": 0.015625,
          "paraphrase_len6_op_accuracy": 0.3177083333333333,
          "paraphrase_len6_arg_accuracy": 0.005208333333333333,
          "paraphrase_len6_program_exact": 0.0,
          "paraphrase_len12_direct_accuracy": NaN,
          "paraphrase_len12_direct_target_mass": NaN,
          "paraphrase_len12_executor_accuracy": 0.0,
          "paraphrase_len12_executor_target_mass": 0.009745899587869644,
          "paraphrase_len12_init_accuracy": 0.03125,
          "paraphrase_len12_op_accuracy": 0.3111979166666667,
          "paraphrase_len12_arg_accuracy": 0.00390625,
          "paraphrase_len12_program_exact": 0.0
        }
      ],
      "final_metrics": {
        "standard_len3": {
          "variant": "compiler_answer_only",
          "n": 64.0,
          "direct_accuracy": NaN,
          "direct_target_mass": NaN,
          "executor_accuracy": 0.03125,
          "executor_target_mass": 0.015396549366414547,
          "init_accuracy": 0.0,
          "op_accuracy": 0.2916666666666667,
          "arg_accuracy": 0.0,
          "program_exact": 0.0
        },
        "standard_len6": {
          "variant": "compiler_answer_only",
          "n": 64.0,
          "direct_accuracy": NaN,
          "direct_target_mass": NaN,
          "executor_accuracy": 0.0,
          "executor_target_mass": 0.010013941675424576,
          "init_accuracy": 0.03125,
          "op_accuracy": 0.3515625,
          "arg_accuracy": 0.0,
          "program_exact": 0.0
        },
        "standard_len12": {
          "variant": "compiler_answer_only",
          "n": 64.0,
          "direct_accuracy": NaN,
          "direct_target_mass": NaN,
          "executor_accuracy": 0.0,
          "executor_target_mass": 0.010086162015795708,
          "init_accuracy": 0.015625,
          "op_accuracy": 0.3177083333333333,
          "arg_accuracy": 0.006510416666666667,
          "program_exact": 0.0
        },
        "paraphrase_len3": {
          "variant": "compiler_answer_only",
          "n": 64.0,
          "direct_accuracy": NaN,
          "direct_target_mass": NaN,
          "executor_accuracy": 0.0,
          "executor_target_mass": 0.006057328544557095,
          "init_accuracy": 0.015625,
          "op_accuracy": 0.3489583333333333,
          "arg_accuracy": 0.0,
          "program_exact": 0.0
        },
        "paraphrase_len6": {
          "variant": "compiler_answer_only",
          "n": 64.0,
          "direct_accuracy": NaN,
          "direct_target_mass": NaN,
          "executor_accuracy": 0.015625,
          "executor_target_mass": 0.008675915189087391,
          "init_accuracy": 0.015625,
          "op_accuracy": 0.3177083333333333,
          "arg_accuracy": 0.005208333333333333,
          "program_exact": 0.0
        },
        "paraphrase_len12": {
          "variant": "compiler_answer_only",
          "n": 64.0,
          "direct_accuracy": NaN,
          "direct_target_mass": NaN,
          "executor_accuracy": 0.0,
          "executor_target_mass": 0.009745899587869644,
          "init_accuracy": 0.03125,
          "op_accuracy": 0.3111979166666667,
          "arg_accuracy": 0.00390625,
          "program_exact": 0.0
        }
      }
    },
    "compiler_trace_then_answer": {
      "variant": "compiler_trace_then_answer",
      "train_seconds": 39.23384141921997,
      "checkpoints": [
        "large_artifacts/qwen_span_free_compiler/checkpoints/pilot_qwen35_span_free/compiler_trace_then_answer/checkpoint_final.pt"
      ],
      "train_log": [
        {
          "stage": "trace_bootstrap",
          "local_step": 1,
          "step": 1,
          "loss": 14.833468437194824,
          "trace_loss_active": true,
          "lr_scale": 1.0,
          "executor_loss": 4.5627851486206055,
          "init_loss": 4.581374645233154,
          "op_loss": 1.0881401300430298,
          "arg_loss": 4.601168632507324,
          "standard_len3_direct_accuracy": NaN,
          "standard_len3_direct_target_mass": NaN,
          "standard_len3_executor_accuracy": 0.0,
          "standard_len3_executor_target_mass": 0.010297341272234917,
          "standard_len3_init_accuracy": 0.0,
          "standard_len3_op_accuracy": 0.3125,
          "standard_len3_arg_accuracy": 0.036458333333333336,
          "standard_len3_program_exact": 0.0,
          "standard_len6_direct_accuracy": NaN,
          "standard_len6_direct_target_mass": NaN,
          "standard_len6_executor_accuracy": 0.0,
          "standard_len6_executor_target_mass": 0.010314024984836578,
          "standard_len6_init_accuracy": 0.0,
          "standard_len6_op_accuracy": 0.2994791666666667,
          "standard_len6_arg_accuracy": 0.052083333333333336,
          "standard_len6_program_exact": 0.0,
          "standard_len12_direct_accuracy": NaN,
          "standard_len12_direct_target_mass": NaN,
          "standard_len12_executor_accuracy": 0.015625,
          "standard_len12_executor_target_mass": 0.010301460511982441,
          "standard_len12_init_accuracy": 0.015625,
          "standard_len12_op_accuracy": 0.3515625,
          "standard_len12_arg_accuracy": 0.046875,
          "standard_len12_program_exact": 0.0,
          "paraphrase_len3_direct_accuracy": NaN,
          "paraphrase_len3_direct_target_mass": NaN,
          "paraphrase_len3_executor_accuracy": 0.046875,
          "paraphrase_len3_executor_target_mass": 0.010315150022506714,
          "paraphrase_len3_init_accuracy": 0.015625,
          "paraphrase_len3_op_accuracy": 0.3385416666666667,
          "paraphrase_len3_arg_accuracy": 0.041666666666666664,
          "paraphrase_len3_program_exact": 0.0,
          "paraphrase_len6_direct_accuracy": NaN,
          "paraphrase_len6_direct_target_mass": NaN,
          "paraphrase_len6_executor_accuracy": 0.0,
          "paraphrase_len6_executor_target_mass": 0.01031368225812912,
          "paraphrase_len6_init_accuracy": 0.03125,
          "paraphrase_len6_op_accuracy": 0.3125,
          "paraphrase_len6_arg_accuracy": 0.036458333333333336,
          "paraphrase_len6_program_exact": 0.0,
          "paraphrase_len12_direct_accuracy": NaN,
          "paraphrase_len12_direct_target_mass": NaN,
          "paraphrase_len12_executor_accuracy": 0.0,
          "paraphrase_len12_executor_target_mass": 0.010300764814019203,
          "paraphrase_len12_init_accuracy": 0.0,
          "paraphrase_len12_op_accuracy": 0.3255208333333333,
          "paraphrase_len12_arg_accuracy": 0.053385416666666664,
          "paraphrase_len12_program_exact": 0.0
        },
        {
          "stage": "trace_bootstrap",
          "local_step": 120,
          "step": 120,
          "loss": 11.815940856933594,
          "trace_loss_active": true,
          "lr_scale": 1.0,
          "executor_loss": 4.562437057495117,
          "init_loss": 3.743434429168701,
          "op_loss": 0.033884137868881226,
          "arg_loss": 3.4761853218078613,
          "standard_len3_direct_accuracy": NaN,
          "standard_len3_direct_target_mass": NaN,
          "standard_len3_executor_accuracy": 0.015625,
          "standard_len3_executor_target_mass": 0.010332669131457806,
          "standard_len3_init_accuracy": 0.015625,
          "standard_len3_op_accuracy": 0.9947916666666666,
          "standard_len3_arg_accuracy": 0.03125,
          "standard_len3_program_exact": 0.0,
          "standard_len6_direct_accuracy": NaN,
          "standard_len6_direct_target_mass": NaN,
          "standard_len6_executor_accuracy": 0.0,
          "standard_len6_executor_target_mass": 0.010321332141757011,
          "standard_len6_init_accuracy": 0.015625,
          "standard_len6_op_accuracy": 0.5729166666666666,
          "standard_len6_arg_accuracy": 0.0546875,
          "standard_len6_program_exact": 0.0,
          "standard_len12_direct_accuracy": NaN,
          "standard_len12_direct_target_mass": NaN,
          "standard_len12_executor_accuracy": 0.015625,
          "standard_len12_executor_target_mass": 0.010306202806532383,
          "standard_len12_init_accuracy": 0.015625,
          "standard_len12_op_accuracy": 0.4752604166666667,
          "standard_len12_arg_accuracy": 0.05859375,
          "standard_len12_program_exact": 0.0,
          "paraphrase_len3_direct_accuracy": NaN,
          "paraphrase_len3_direct_target_mass": NaN,
          "paraphrase_len3_executor_accuracy": 0.0,
          "paraphrase_len3_executor_target_mass": 0.01049754023551941,
          "paraphrase_len3_init_accuracy": 0.0,
          "paraphrase_len3_op_accuracy": 0.4895833333333333,
          "paraphrase_len3_arg_accuracy": 0.052083333333333336,
          "paraphrase_len3_program_exact": 0.0,
          "paraphrase_len6_direct_accuracy": NaN,
          "paraphrase_len6_direct_target_mass": NaN,
          "paraphrase_len6_executor_accuracy": 0.015625,
          "paraphrase_len6_executor_target_mass": 0.010314682498574257,
          "paraphrase_len6_init_accuracy": 0.0,
          "paraphrase_len6_op_accuracy": 0.3880208333333333,
          "paraphrase_len6_arg_accuracy": 0.057291666666666664,
          "paraphrase_len6_program_exact": 0.0,
          "paraphrase_len12_direct_accuracy": NaN,
          "paraphrase_len12_direct_target_mass": NaN,
          "paraphrase_len12_executor_accuracy": 0.0,
          "paraphrase_len12_executor_target_mass": 0.010309494100511074,
          "paraphrase_len12_init_accuracy": 0.046875,
          "paraphrase_len12_op_accuracy": 0.3854166666666667,
          "paraphrase_len12_arg_accuracy": 0.0390625,
          "paraphrase_len12_program_exact": 0.0
        },
        {
          "stage": "trace_bootstrap",
          "local_step": 240,
          "step": 240,
          "loss": 10.60389518737793,
          "trace_loss_active": true,
          "lr_scale": 1.0,
          "executor_loss": 4.387653350830078,
          "init_loss": 3.018942356109619,
          "op_loss": 0.001453853095881641,
          "arg_loss": 3.1958460807800293,
          "standard_len3_direct_accuracy": NaN,
          "standard_len3_direct_target_mass": NaN,
          "standard_len3_executor_accuracy": 0.0625,
          "standard_len3_executor_target_mass": 0.010224180296063423,
          "standard_len3_init_accuracy": 0.015625,
          "standard_len3_op_accuracy": 0.9895833333333334,
          "standard_len3_arg_accuracy": 0.046875,
          "standard_len3_program_exact": 0.0,
          "standard_len6_direct_accuracy": NaN,
          "standard_len6_direct_target_mass": NaN,
          "standard_len6_executor_accuracy": 0.0,
          "standard_len6_executor_target_mass": 0.010279480367898941,
          "standard_len6_init_accuracy": 0.0,
          "standard_len6_op_accuracy": 0.578125,
          "standard_len6_arg_accuracy": 0.057291666666666664,
          "standard_len6_program_exact": 0.0,
          "standard_len12_direct_accuracy": NaN,
          "standard_len12_direct_target_mass": NaN,
          "standard_len12_executor_accuracy": 0.0,
          "standard_len12_executor_target_mass": 0.010310960933566093,
          "standard_len12_init_accuracy": 0.046875,
          "standard_len12_op_accuracy": 0.4427083333333333,
          "standard_len12_arg_accuracy": 0.0546875,
          "standard_len12_program_exact": 0.0,
          "paraphrase_len3_direct_accuracy": NaN,
          "paraphrase_len3_direct_target_mass": NaN,
          "paraphrase_len3_executor_accuracy": 0.015625,
          "paraphrase_len3_executor_target_mass": 0.010311946272850037,
          "paraphrase_len3_init_accuracy": 0.015625,
          "paraphrase_len3_op_accuracy": 0.5052083333333334,
          "paraphrase_len3_arg_accuracy": 0.041666666666666664,
          "paraphrase_len3_program_exact": 0.0,
          "paraphrase_len6_direct_accuracy": NaN,
          "paraphrase_len6_direct_target_mass": NaN,
          "paraphrase_len6_executor_accuracy": 0.015625,
          "paraphrase_len6_executor_target_mass": 0.010524173267185688,
          "paraphrase_len6_init_accuracy": 0.0,
          "paraphrase_len6_op_accuracy": 0.3932291666666667,
          "paraphrase_len6_arg_accuracy": 0.033854166666666664,
          "paraphrase_len6_program_exact": 0.0,
          "paraphrase_len12_direct_accuracy": NaN,
          "paraphrase_len12_direct_target_mass": NaN,
          "paraphrase_len12_executor_accuracy": 0.0,
          "paraphrase_len12_executor_target_mass": 0.010279817506670952,
          "paraphrase_len12_init_accuracy": 0.015625,
          "paraphrase_len12_op_accuracy": 0.3932291666666667,
          "paraphrase_len12_arg_accuracy": 0.057291666666666664,
          "paraphrase_len12_program_exact": 0.0
        },
        {
          "stage": "answer_retention",
          "local_step": 1,
          "step": 241,
          "loss": 4.663530349731445,
          "trace_loss_active": false,
          "lr_scale": 1.0,
          "executor_loss": 4.663530349731445,
          "standard_len3_direct_accuracy": NaN,
          "standard_len3_direct_target_mass": NaN,
          "standard_len3_executor_accuracy": 0.03125,
          "standard_len3_executor_target_mass": 0.010289536789059639,
          "standard_len3_init_accuracy": 0.015625,
          "standard_len3_op_accuracy": 0.9947916666666666,
          "standard_len3_arg_accuracy": 0.036458333333333336,
          "standard_len3_program_exact": 0.0,
          "standard_len6_direct_accuracy": NaN,
          "standard_len6_direct_target_mass": NaN,
          "standard_len6_executor_accuracy": 0.0,
          "standard_len6_executor_target_mass": 0.010334662161767483,
          "standard_len6_init_accuracy": 0.0,
          "standard_len6_op_accuracy": 0.5729166666666666,
          "standard_len6_arg_accuracy": 0.041666666666666664,
          "standard_len6_program_exact": 0.0,
          "standard_len12_direct_accuracy": NaN,
          "standard_len12_direct_target_mass": NaN,
          "standard_len12_executor_accuracy": 0.015625,
          "standard_len12_executor_target_mass": 0.010309537872672081,
          "standard_len12_init_accuracy": 0.015625,
          "standard_len12_op_accuracy": 0.4427083333333333,
          "standard_len12_arg_accuracy": 0.049479166666666664,
          "standard_len12_program_exact": 0.0,
          "paraphrase_len3_direct_accuracy": NaN,
          "paraphrase_len3_direct_target_mass": NaN,
          "paraphrase_len3_executor_accuracy": 0.015625,
          "paraphrase_len3_executor_target_mass": 0.010261311195790768,
          "paraphrase_len3_init_accuracy": 0.015625,
          "paraphrase_len3_op_accuracy": 0.515625,
          "paraphrase_len3_arg_accuracy": 0.041666666666666664,
          "paraphrase_len3_program_exact": 0.0,
          "paraphrase_len6_direct_accuracy": NaN,
          "paraphrase_len6_direct_target_mass": NaN,
          "paraphrase_len6_executor_accuracy": 0.015625,
          "paraphrase_len6_executor_target_mass": 0.010403030551970005,
          "paraphrase_len6_init_accuracy": 0.015625,
          "paraphrase_len6_op_accuracy": 0.4036458333333333,
          "paraphrase_len6_arg_accuracy": 0.036458333333333336,
          "paraphrase_len6_program_exact": 0.0,
          "paraphrase_len12_direct_accuracy": NaN,
          "paraphrase_len12_direct_target_mass": NaN,
          "paraphrase_len12_executor_accuracy": 0.0,
          "paraphrase_len12_executor_target_mass": 0.01028364896774292,
          "paraphrase_len12_init_accuracy": 0.015625,
          "paraphrase_len12_op_accuracy": 0.38671875,
          "paraphrase_len12_arg_accuracy": 0.055989583333333336,
          "paraphrase_len12_program_exact": 0.0
        },
        {
          "stage": "answer_retention",
          "local_step": 120,
          "step": 360,
          "loss": 4.278226852416992,
          "trace_loss_active": false,
          "lr_scale": 1.0,
          "executor_loss": 4.278226852416992,
          "standard_len3_direct_accuracy": NaN,
          "standard_len3_direct_target_mass": NaN,
          "standard_len3_executor_accuracy": 0.0,
          "standard_len3_executor_target_mass": 0.008534278720617294,
          "standard_len3_init_accuracy": 0.03125,
          "standard_len3_op_accuracy": 0.8072916666666666,
          "standard_len3_arg_accuracy": 0.020833333333333332,
          "standard_len3_program_exact": 0.0,
          "standard_len6_direct_accuracy": NaN,
          "standard_len6_direct_target_mass": NaN,
          "standard_len6_executor_accuracy": 0.0,
          "standard_len6_executor_target_mass": 0.010036060586571693,
          "standard_len6_init_accuracy": 0.0,
          "standard_len6_op_accuracy": 0.5104166666666666,
          "standard_len6_arg_accuracy": 0.018229166666666668,
          "standard_len6_program_exact": 0.0,
          "standard_len12_direct_accuracy": NaN,
          "standard_len12_direct_target_mass": NaN,
          "standard_len12_executor_accuracy": 0.015625,
          "standard_len12_executor_target_mass": 0.010303874500095844,
          "standard_len12_init_accuracy": 0.03125,
          "standard_len12_op_accuracy": 0.4453125,
          "standard_len12_arg_accuracy": 0.022135416666666668,
          "standard_len12_program_exact": 0.0,
          "paraphrase_len3_direct_accuracy": NaN,
          "paraphrase_len3_direct_target_mass": NaN,
          "paraphrase_len3_executor_accuracy": 0.0,
          "paraphrase_len3_executor_target_mass": 0.010141192004084587,
          "paraphrase_len3_init_accuracy": 0.03125,
          "paraphrase_len3_op_accuracy": 0.4479166666666667,
          "paraphrase_len3_arg_accuracy": 0.03125,
          "paraphrase_len3_program_exact": 0.0,
          "paraphrase_len6_direct_accuracy": NaN,
          "paraphrase_len6_direct_target_mass": NaN,
          "paraphrase_len6_executor_accuracy": 0.015625,
          "paraphrase_len6_executor_target_mass": 0.010182695463299751,
          "paraphrase_len6_init_accuracy": 0.015625,
          "paraphrase_len6_op_accuracy": 0.3776041666666667,
          "paraphrase_len6_arg_accuracy": 0.018229166666666668,
          "paraphrase_len6_program_exact": 0.0,
          "paraphrase_len12_direct_accuracy": NaN,
          "paraphrase_len12_direct_target_mass": NaN,
          "paraphrase_len12_executor_accuracy": 0.0,
          "paraphrase_len12_executor_target_mass": 0.010307811200618744,
          "paraphrase_len12_init_accuracy": 0.0,
          "paraphrase_len12_op_accuracy": 0.3671875,
          "paraphrase_len12_arg_accuracy": 0.018229166666666668,
          "paraphrase_len12_program_exact": 0.0
        },
        {
          "stage": "answer_retention",
          "local_step": 240,
          "step": 480,
          "loss": 3.687770366668701,
          "trace_loss_active": false,
          "lr_scale": 1.0,
          "executor_loss": 3.687770366668701,
          "standard_len3_direct_accuracy": NaN,
          "standard_len3_direct_target_mass": NaN,
          "standard_len3_executor_accuracy": 0.0,
          "standard_len3_executor_target_mass": 0.006027504801750183,
          "standard_len3_init_accuracy": 0.015625,
          "standard_len3_op_accuracy": 0.8177083333333334,
          "standard_len3_arg_accuracy": 0.005208333333333333,
          "standard_len3_program_exact": 0.0,
          "standard_len6_direct_accuracy": NaN,
          "standard_len6_direct_target_mass": NaN,
          "standard_len6_executor_accuracy": 0.0,
          "standard_len6_executor_target_mass": 0.009071755222976208,
          "standard_len6_init_accuracy": 0.015625,
          "standard_len6_op_accuracy": 0.5104166666666666,
          "standard_len6_arg_accuracy": 0.020833333333333332,
          "standard_len6_program_exact": 0.0,
          "standard_len12_direct_accuracy": NaN,
          "standard_len12_direct_target_mass": NaN,
          "standard_len12_executor_accuracy": 0.015625,
          "standard_len12_executor_target_mass": 0.010918930172920227,
          "standard_len12_init_accuracy": 0.03125,
          "standard_len12_op_accuracy": 0.4244791666666667,
          "standard_len12_arg_accuracy": 0.020833333333333332,
          "standard_len12_program_exact": 0.0,
          "paraphrase_len3_direct_accuracy": NaN,
          "paraphrase_len3_direct_target_mass": NaN,
          "paraphrase_len3_executor_accuracy": 0.0,
          "paraphrase_len3_executor_target_mass": 0.009516744874417782,
          "paraphrase_len3_init_accuracy": 0.015625,
          "paraphrase_len3_op_accuracy": 0.4270833333333333,
          "paraphrase_len3_arg_accuracy": 0.041666666666666664,
          "paraphrase_len3_program_exact": 0.0,
          "paraphrase_len6_direct_accuracy": NaN,
          "paraphrase_len6_direct_target_mass": NaN,
          "paraphrase_len6_executor_accuracy": 0.03125,
          "paraphrase_len6_executor_target_mass": 0.01027028076350689,
          "paraphrase_len6_init_accuracy": 0.0,
          "paraphrase_len6_op_accuracy": 0.3645833333333333,
          "paraphrase_len6_arg_accuracy": 0.015625,
          "paraphrase_len6_program_exact": 0.0,
          "paraphrase_len12_direct_accuracy": NaN,
          "paraphrase_len12_direct_target_mass": NaN,
          "paraphrase_len12_executor_accuracy": 0.015625,
          "paraphrase_len12_executor_target_mass": 0.010310983285307884,
          "paraphrase_len12_init_accuracy": 0.0,
          "paraphrase_len12_op_accuracy": 0.3697916666666667,
          "paraphrase_len12_arg_accuracy": 0.029947916666666668,
          "paraphrase_len12_program_exact": 0.0
        }
      ],
      "final_metrics": {
        "standard_len3": {
          "variant": "compiler_trace_then_answer",
          "n": 64.0,
          "direct_accuracy": NaN,
          "direct_target_mass": NaN,
          "executor_accuracy": 0.0,
          "executor_target_mass": 0.006027504801750183,
          "init_accuracy": 0.015625,
          "op_accuracy": 0.8177083333333334,
          "arg_accuracy": 0.005208333333333333,
          "program_exact": 0.0
        },
        "standard_len6": {
          "variant": "compiler_trace_then_answer",
          "n": 64.0,
          "direct_accuracy": NaN,
          "direct_target_mass": NaN,
          "executor_accuracy": 0.0,
          "executor_target_mass": 0.009071755222976208,
          "init_accuracy": 0.015625,
          "op_accuracy": 0.5104166666666666,
          "arg_accuracy": 0.020833333333333332,
          "program_exact": 0.0
        },
        "standard_len12": {
          "variant": "compiler_trace_then_answer",
          "n": 64.0,
          "direct_accuracy": NaN,
          "direct_target_mass": NaN,
          "executor_accuracy": 0.015625,
          "executor_target_mass": 0.010918930172920227,
          "init_accuracy": 0.03125,
          "op_accuracy": 0.4244791666666667,
          "arg_accuracy": 0.020833333333333332,
          "program_exact": 0.0
        },
        "paraphrase_len3": {
          "variant": "compiler_trace_then_answer",
          "n": 64.0,
          "direct_accuracy": NaN,
          "direct_target_mass": NaN,
          "executor_accuracy": 0.0,
          "executor_target_mass": 0.009516744874417782,
          "init_accuracy": 0.015625,
          "op_accuracy": 0.4270833333333333,
          "arg_accuracy": 0.041666666666666664,
          "program_exact": 0.0
        },
        "paraphrase_len6": {
          "variant": "compiler_trace_then_answer",
          "n": 64.0,
          "direct_accuracy": NaN,
          "direct_target_mass": NaN,
          "executor_accuracy": 0.03125,
          "executor_target_mass": 0.01027028076350689,
          "init_accuracy": 0.0,
          "op_accuracy": 0.3645833333333333,
          "arg_accuracy": 0.015625,
          "program_exact": 0.0
        },
        "paraphrase_len12": {
          "variant": "compiler_trace_then_answer",
          "n": 64.0,
          "direct_accuracy": NaN,
          "direct_target_mass": NaN,
          "executor_accuracy": 0.015625,
          "executor_target_mass": 0.010310983285307884,
          "init_accuracy": 0.0,
          "op_accuracy": 0.3697916666666667,
          "arg_accuracy": 0.029947916666666668,
          "program_exact": 0.0
        }
      }
    }
  }
}