{
  "run": "pilot_expand_copy",
  "arm": "expand_copy",
  "args": {
    "run_name": "pilot_expand_copy",
    "arm": "expand_copy",
    "model_id": "Qwen/Qwen3-4B",
    "seed": 123,
    "modulus": 97,
    "stage_max_steps": "8,16,24",
    "stage_min_lengths": "1,1,8",
    "stage_train_max_lengths": "8,16,24",
    "stage_steps": "30,15,30",
    "train_examples": 96,
    "train_template_mode": "mixed",
    "eval_lengths": "8,16,24",
    "eval_examples": 8,
    "paired_eval_pairs": 4,
    "batch_size": 4,
    "grad_accum": 1,
    "eval_batch_size": 4,
    "max_length": 2048,
    "head_width": 256,
    "compiler_layers": 1,
    "compiler_heads": 4,
    "compiler_dropout": 0.05,
    "expansion_noise": 0.005,
    "expansion_mode": "copy_last",
    "trace_loss_weight": 1.0,
    "executor_loss_weight": 1.0,
    "state_loss_weight": 1.0,
    "init_trace_loss_weight": 4.0,
    "op_trace_loss_weight": 1.0,
    "arg_trace_loss_weight": 4.0,
    "direct_head_weight": 0.0,
    "lr": 0.0002,
    "weight_decay": 0.0,
    "optimizer": "paged_adamw_8bit",
    "max_grad_norm": 1.0,
    "torch_dtype": "bf16",
    "load_in_4bit": true,
    "device_map": "auto",
    "use_lora": true,
    "gradient_checkpointing": true,
    "lora_r": 8,
    "lora_alpha": 16,
    "lora_dropout": 0.05,
    "lora_target_modules": "all-linear",
    "log_interval": 15,
    "save_checkpoints": false,
    "checkpoint_dir": "",
    "selection_split": "paired_L24",
    "selection_metric": "executor_pair_both_correct"
  },
  "metadata": {
    "python": "3.12.3",
    "platform": "Linux-6.8.0-62-generic-x86_64-with-glibc2.39",
    "torch": "2.8.0+cu128",
    "cuda_available": true,
    "transformers_loader": "AutoModelForCausalLM",
    "peft_installed": true,
    "use_lora": true,
    "lora_r": 8,
    "model_id": "Qwen/Qwen3-4B",
    "gpu_name": "NVIDIA RTX 6000 Ada Generation",
    "gpu_vram_gb": 47.383
  },
  "stages": [
    {
      "name": "stage1_max8",
      "max_steps": 8,
      "train_min_len": 1,
      "train_max_len": 8,
      "steps": 30
    },
    {
      "name": "stage2_max16",
      "max_steps": 16,
      "train_min_len": 1,
      "train_max_len": 16,
      "steps": 15
    },
    {
      "name": "stage3_max24",
      "max_steps": 24,
      "train_min_len": 8,
      "train_max_len": 24,
      "steps": 30
    }
  ],
  "train_log": [
    {
      "run": "pilot_expand_copy",
      "arm": "expand_copy",
      "stage": "stage1_max8",
      "stage_idx": 1,
      "local_step": 1,
      "global_step": 1,
      "compiler_max_steps": 8,
      "elapsed_sec": 1.327,
      "expansion_event": "init",
      "init_loss": 4.4699859619140625,
      "op_loss": 1.1295549869537354,
      "arg_loss": 4.528594017028809,
      "state_loss": 4.580868721008301,
      "state_train_accuracy": 0.0,
      "executor_loss": 4.580234527587891,
      "loss": 46.284976959228516
    },
    {
      "run": "pilot_expand_copy",
      "arm": "expand_copy",
      "stage": "stage1_max8",
      "stage_idx": 1,
      "local_step": 15,
      "global_step": 15,
      "compiler_max_steps": 8,
      "elapsed_sec": 8.506,
      "expansion_event": "init",
      "init_loss": 4.678859710693359,
      "op_loss": 1.0444238185882568,
      "arg_loss": 4.231142520904541,
      "state_loss": 4.576465606689453,
      "state_train_accuracy": 0.0,
      "executor_loss": 4.57564640045166,
      "loss": 45.836544036865234
    },
    {
      "run": "pilot_expand_copy",
      "arm": "expand_copy",
      "stage": "stage1_max8",
      "stage_idx": 1,
      "local_step": 30,
      "global_step": 30,
      "compiler_max_steps": 8,
      "elapsed_sec": 16.286,
      "expansion_event": "init",
      "init_loss": 4.800575256347656,
      "op_loss": 1.2056492567062378,
      "arg_loss": 3.817878007888794,
      "state_loss": 4.575311660766602,
      "state_train_accuracy": 0.0,
      "executor_loss": 4.577898025512695,
      "loss": 44.832672119140625
    },
    {
      "run": "pilot_expand_copy",
      "arm": "expand_copy",
      "stage": "stage2_max16",
      "stage_idx": 2,
      "local_step": 1,
      "global_step": 31,
      "compiler_max_steps": 16,
      "elapsed_sec": 19.547,
      "expansion_event": "expand_8_to_16_copy_last",
      "init_loss": 5.027321815490723,
      "op_loss": 1.0847711563110352,
      "arg_loss": 3.9461615085601807,
      "state_loss": 4.576554298400879,
      "state_train_accuracy": 0.0,
      "executor_loss": 4.576953887939453,
      "loss": 46.13220977783203
    },
    {
      "run": "pilot_expand_copy",
      "arm": "expand_copy",
      "stage": "stage2_max16",
      "stage_idx": 2,
      "local_step": 15,
      "global_step": 45,
      "compiler_max_steps": 16,
      "elapsed_sec": 34.054,
      "expansion_event": "expand_8_to_16_copy_last",
      "init_loss": 4.383572578430176,
      "op_loss": 1.0575400590896606,
      "arg_loss": 3.703749418258667,
      "state_loss": 4.579343318939209,
      "state_train_accuracy": 0.0,
      "executor_loss": 4.581912994384766,
      "loss": 42.568084716796875
    },
    {
      "run": "pilot_expand_copy",
      "arm": "expand_copy",
      "stage": "stage3_max24",
      "stage_idx": 3,
      "local_step": 1,
      "global_step": 46,
      "compiler_max_steps": 24,
      "elapsed_sec": 43.137,
      "expansion_event": "expand_16_to_24_copy_last",
      "init_loss": 4.74779748916626,
      "op_loss": 1.1165481805801392,
      "arg_loss": 3.7822458744049072,
      "state_loss": 4.578446388244629,
      "state_train_accuracy": 0.0,
      "executor_loss": 4.579893112182617,
      "loss": 44.395057678222656
    },
    {
      "run": "pilot_expand_copy",
      "arm": "expand_copy",
      "stage": "stage3_max24",
      "stage_idx": 3,
      "local_step": 15,
      "global_step": 60,
      "compiler_max_steps": 24,
      "elapsed_sec": 65.511,
      "expansion_event": "expand_16_to_24_copy_last",
      "init_loss": 4.8124003410339355,
      "op_loss": 1.120258092880249,
      "arg_loss": 3.726202964782715,
      "state_loss": 4.577245712280273,
      "state_train_accuracy": 0.0,
      "executor_loss": 4.577106475830078,
      "loss": 44.42902374267578
    },
    {
      "run": "pilot_expand_copy",
      "arm": "expand_copy",
      "stage": "stage3_max24",
      "stage_idx": 3,
      "local_step": 30,
      "global_step": 75,
      "compiler_max_steps": 24,
      "elapsed_sec": 90.205,
      "expansion_event": "expand_16_to_24_copy_last",
      "init_loss": 3.4682273864746094,
      "op_loss": 1.0923902988433838,
      "arg_loss": 3.7079877853393555,
      "state_loss": 4.574213981628418,
      "state_train_accuracy": 0.0,
      "executor_loss": 4.575571060180664,
      "loss": 38.94703674316406
    }
  ],
  "metrics": [
    {
      "stage": "stage1_max8",
      "split": "standard_L8",
      "n": 8,
      "max_steps": 8,
      "executor_accuracy": 0.0,
      "executor_target_mass": 0.010268445126712322,
      "direct_accuracy": NaN,
      "init_accuracy": 0.0,
      "op_accuracy": 0.328125,
      "arg_accuracy": 0.09375,
      "program_exact": 0.0,
      "state_accuracy": 0.0,
      "state_all_exact": 0.0,
      "state_prefix_fraction": 0.0,
      "state_mean_correct_prefix": 0.0,
      "run": "pilot_expand_copy",
      "arm": "expand_copy",
      "global_step": 30,
      "stage_idx": 1
    },
    {
      "stage": "stage1_max8",
      "split": "paraphrase_L8",
      "n": 8,
      "max_steps": 8,
      "executor_accuracy": 0.0,
      "executor_target_mass": 0.010269857943058014,
      "direct_accuracy": NaN,
      "init_accuracy": 0.0,
      "op_accuracy": 0.375,
      "arg_accuracy": 0.0625,
      "program_exact": 0.0,
      "state_accuracy": 0.0,
      "state_all_exact": 0.0,
      "state_prefix_fraction": 0.0,
      "state_mean_correct_prefix": 0.0,
      "run": "pilot_expand_copy",
      "arm": "expand_copy",
      "global_step": 30,
      "stage_idx": 1
    },
    {
      "stage": "stage1_max8",
      "split": "heldout_L8",
      "n": 8,
      "max_steps": 8,
      "executor_accuracy": 0.0,
      "executor_target_mass": 0.010268381796777248,
      "direct_accuracy": NaN,
      "init_accuracy": 0.0,
      "op_accuracy": 0.34375,
      "arg_accuracy": 0.046875,
      "program_exact": 0.0,
      "state_accuracy": 0.0,
      "state_all_exact": 0.0,
      "state_prefix_fraction": 0.0,
      "state_mean_correct_prefix": 0.0,
      "run": "pilot_expand_copy",
      "arm": "expand_copy",
      "global_step": 30,
      "stage_idx": 1
    },
    {
      "stage": "stage1_max8",
      "split": "paired_L8",
      "n": 8,
      "max_steps": 8,
      "executor_accuracy": 0.0,
      "executor_target_mass": 0.010279937647283077,
      "direct_accuracy": NaN,
      "init_accuracy": 0.0,
      "op_accuracy": 0.3125,
      "arg_accuracy": 0.0625,
      "program_exact": 0.0,
      "state_accuracy": 0.0,
      "state_all_exact": 0.0,
      "state_prefix_fraction": 0.0,
      "state_mean_correct_prefix": 0.0,
      "executor_pair_answer_consistency": 1.0,
      "executor_pair_both_correct": 0.0,
      "compiler_pair_program_consistency": 1.0,
      "compiler_pair_state_consistency": 1.0,
      "run": "pilot_expand_copy",
      "arm": "expand_copy",
      "global_step": 30,
      "stage_idx": 1
    },
    {
      "stage": "stage1_max8",
      "split": "paired_heldout_L8",
      "n": 8,
      "max_steps": 8,
      "executor_accuracy": 0.25,
      "executor_target_mass": 0.010263872798532248,
      "direct_accuracy": NaN,
      "init_accuracy": 0.0,
      "op_accuracy": 0.4375,
      "arg_accuracy": 0.0,
      "program_exact": 0.0,
      "state_accuracy": 0.03125,
      "state_all_exact": 0.0,
      "state_prefix_fraction": 0.0,
      "state_mean_correct_prefix": 0.0,
      "executor_pair_answer_consistency": 1.0,
      "executor_pair_both_correct": 0.25,
      "compiler_pair_program_consistency": 1.0,
      "compiler_pair_state_consistency": 1.0,
      "run": "pilot_expand_copy",
      "arm": "expand_copy",
      "global_step": 30,
      "stage_idx": 1
    },
    {
      "stage": "stage2_max16",
      "split": "standard_L8",
      "n": 8,
      "max_steps": 16,
      "executor_accuracy": 0.0,
      "executor_target_mass": 0.010767539963126183,
      "direct_accuracy": NaN,
      "init_accuracy": 0.0,
      "op_accuracy": 0.296875,
      "arg_accuracy": 0.046875,
      "program_exact": 0.0,
      "state_accuracy": 0.0,
      "state_all_exact": 0.0,
      "state_prefix_fraction": 0.0,
      "state_mean_correct_prefix": 0.0,
      "run": "pilot_expand_copy",
      "arm": "expand_copy",
      "global_step": 45,
      "stage_idx": 2
    },
    {
      "stage": "stage2_max16",
      "split": "paraphrase_L8",
      "n": 8,
      "max_steps": 16,
      "executor_accuracy": 0.125,
      "executor_target_mass": 0.010276052169501781,
      "direct_accuracy": NaN,
      "init_accuracy": 0.0,
      "op_accuracy": 0.40625,
      "arg_accuracy": 0.015625,
      "program_exact": 0.0,
      "state_accuracy": 0.03125,
      "state_all_exact": 0.0,
      "state_prefix_fraction": 0.0,
      "state_mean_correct_prefix": 0.0,
      "run": "pilot_expand_copy",
      "arm": "expand_copy",
      "global_step": 45,
      "stage_idx": 2
    },
    {
      "stage": "stage2_max16",
      "split": "heldout_L8",
      "n": 8,
      "max_steps": 16,
      "executor_accuracy": 0.0,
      "executor_target_mass": 0.010268719866871834,
      "direct_accuracy": NaN,
      "init_accuracy": 0.0,
      "op_accuracy": 0.328125,
      "arg_accuracy": 0.046875,
      "program_exact": 0.0,
      "state_accuracy": 0.0,
      "state_all_exact": 0.0,
      "state_prefix_fraction": 0.0,
      "state_mean_correct_prefix": 0.0,
      "run": "pilot_expand_copy",
      "arm": "expand_copy",
      "global_step": 45,
      "stage_idx": 2
    },
    {
      "stage": "stage2_max16",
      "split": "paired_L8",
      "n": 8,
      "max_steps": 16,
      "executor_accuracy": 0.0,
      "executor_target_mass": 0.01026687165722251,
      "direct_accuracy": NaN,
      "init_accuracy": 0.0,
      "op_accuracy": 0.359375,
      "arg_accuracy": 0.03125,
      "program_exact": 0.0,
      "state_accuracy": 0.0,
      "state_all_exact": 0.0,
      "state_prefix_fraction": 0.0,
      "state_mean_correct_prefix": 0.0,
      "executor_pair_answer_consistency": 0.75,
      "executor_pair_both_correct": 0.0,
      "compiler_pair_program_consistency": 0.75,
      "compiler_pair_state_consistency": 0.75,
      "run": "pilot_expand_copy",
      "arm": "expand_copy",
      "global_step": 45,
      "stage_idx": 2
    },
    {
      "stage": "stage2_max16",
      "split": "paired_heldout_L8",
      "n": 8,
      "max_steps": 16,
      "executor_accuracy": 0.0,
      "executor_target_mass": 0.010265734046697617,
      "direct_accuracy": NaN,
      "init_accuracy": 0.0,
      "op_accuracy": 0.3125,
      "arg_accuracy": 0.0,
      "program_exact": 0.0,
      "state_accuracy": 0.0,
      "state_all_exact": 0.0,
      "state_prefix_fraction": 0.0,
      "state_mean_correct_prefix": 0.0,
      "executor_pair_answer_consistency": 0.75,
      "executor_pair_both_correct": 0.0,
      "compiler_pair_program_consistency": 0.75,
      "compiler_pair_state_consistency": 0.75,
      "run": "pilot_expand_copy",
      "arm": "expand_copy",
      "global_step": 45,
      "stage_idx": 2
    },
    {
      "stage": "stage2_max16",
      "split": "standard_L16",
      "n": 8,
      "max_steps": 16,
      "executor_accuracy": 0.0,
      "executor_target_mass": 0.010262238793075085,
      "direct_accuracy": NaN,
      "init_accuracy": 0.0,
      "op_accuracy": 0.296875,
      "arg_accuracy": 0.0390625,
      "program_exact": 0.0,
      "state_accuracy": 0.0078125,
      "state_all_exact": 0.0,
      "state_prefix_fraction": 0.0,
      "state_mean_correct_prefix": 0.0,
      "run": "pilot_expand_copy",
      "arm": "expand_copy",
      "global_step": 45,
      "stage_idx": 2
    },
    {
      "stage": "stage2_max16",
      "split": "paraphrase_L16",
      "n": 8,
      "max_steps": 16,
      "executor_accuracy": 0.0,
      "executor_target_mass": 0.010258786845952272,
      "direct_accuracy": NaN,
      "init_accuracy": 0.0,
      "op_accuracy": 0.3671875,
      "arg_accuracy": 0.0390625,
      "program_exact": 0.0,
      "state_accuracy": 0.0078125,
      "state_all_exact": 0.0,
      "state_prefix_fraction": 0.0,
      "state_mean_correct_prefix": 0.0,
      "run": "pilot_expand_copy",
      "arm": "expand_copy",
      "global_step": 45,
      "stage_idx": 2
    },
    {
      "stage": "stage2_max16",
      "split": "heldout_L16",
      "n": 8,
      "max_steps": 16,
      "executor_accuracy": 0.125,
      "executor_target_mass": 0.010261962190270424,
      "direct_accuracy": NaN,
      "init_accuracy": 0.0,
      "op_accuracy": 0.359375,
      "arg_accuracy": 0.078125,
      "program_exact": 0.0,
      "state_accuracy": 0.0078125,
      "state_all_exact": 0.0,
      "state_prefix_fraction": 0.0,
      "state_mean_correct_prefix": 0.0,
      "run": "pilot_expand_copy",
      "arm": "expand_copy",
      "global_step": 45,
      "stage_idx": 2
    },
    {
      "stage": "stage2_max16",
      "split": "paired_L16",
      "n": 8,
      "max_steps": 16,
      "executor_accuracy": 0.0,
      "executor_target_mass": 0.010265768505632877,
      "direct_accuracy": NaN,
      "init_accuracy": 0.0,
      "op_accuracy": 0.34375,
      "arg_accuracy": 0.078125,
      "program_exact": 0.0,
      "state_accuracy": 0.015625,
      "state_all_exact": 0.0,
      "state_prefix_fraction": 0.0,
      "state_mean_correct_prefix": 0.0,
      "executor_pair_answer_consistency": 0.75,
      "executor_pair_both_correct": 0.0,
      "compiler_pair_program_consistency": 0.75,
      "compiler_pair_state_consistency": 0.75,
      "run": "pilot_expand_copy",
      "arm": "expand_copy",
      "global_step": 45,
      "stage_idx": 2
    },
    {
      "stage": "stage2_max16",
      "split": "paired_heldout_L16",
      "n": 8,
      "max_steps": 16,
      "executor_accuracy": 0.0,
      "executor_target_mass": 0.010263592936098576,
      "direct_accuracy": NaN,
      "init_accuracy": 0.0,
      "op_accuracy": 0.34375,
      "arg_accuracy": 0.046875,
      "program_exact": 0.0,
      "state_accuracy": 0.0078125,
      "state_all_exact": 0.0,
      "state_prefix_fraction": 0.0,
      "state_mean_correct_prefix": 0.0,
      "executor_pair_answer_consistency": 0.75,
      "executor_pair_both_correct": 0.0,
      "compiler_pair_program_consistency": 0.75,
      "compiler_pair_state_consistency": 0.75,
      "run": "pilot_expand_copy",
      "arm": "expand_copy",
      "global_step": 45,
      "stage_idx": 2
    },
    {
      "stage": "stage3_max24",
      "split": "standard_L8",
      "n": 8,
      "max_steps": 24,
      "executor_accuracy": 0.0,
      "executor_target_mass": 0.010296478401869535,
      "direct_accuracy": NaN,
      "init_accuracy": 0.25,
      "op_accuracy": 0.359375,
      "arg_accuracy": 0.046875,
      "program_exact": 0.0,
      "state_accuracy": 0.0,
      "state_all_exact": 0.0,
      "state_prefix_fraction": 0.0,
      "state_mean_correct_prefix": 0.0,
      "run": "pilot_expand_copy",
      "arm": "expand_copy",
      "global_step": 75,
      "stage_idx": 3
    },
    {
      "stage": "stage3_max24",
      "split": "paraphrase_L8",
      "n": 8,
      "max_steps": 24,
      "executor_accuracy": 0.0,
      "executor_target_mass": 0.01029987819492817,
      "direct_accuracy": NaN,
      "init_accuracy": 0.25,
      "op_accuracy": 0.359375,
      "arg_accuracy": 0.046875,
      "program_exact": 0.0,
      "state_accuracy": 0.0,
      "state_all_exact": 0.0,
      "state_prefix_fraction": 0.0,
      "state_mean_correct_prefix": 0.0,
      "run": "pilot_expand_copy",
      "arm": "expand_copy",
      "global_step": 75,
      "stage_idx": 3
    },
    {
      "stage": "stage3_max24",
      "split": "heldout_L8",
      "n": 8,
      "max_steps": 24,
      "executor_accuracy": 0.0,
      "executor_target_mass": 0.01030312106013298,
      "direct_accuracy": NaN,
      "init_accuracy": 0.0,
      "op_accuracy": 0.328125,
      "arg_accuracy": 0.0625,
      "program_exact": 0.0,
      "state_accuracy": 0.0,
      "state_all_exact": 0.0,
      "state_prefix_fraction": 0.0,
      "state_mean_correct_prefix": 0.0,
      "run": "pilot_expand_copy",
      "arm": "expand_copy",
      "global_step": 75,
      "stage_idx": 3
    },
    {
      "stage": "stage3_max24",
      "split": "paired_L8",
      "n": 8,
      "max_steps": 24,
      "executor_accuracy": 0.0,
      "executor_target_mass": 0.010306192561984062,
      "direct_accuracy": NaN,
      "init_accuracy": 0.0,
      "op_accuracy": 0.25,
      "arg_accuracy": 0.0,
      "program_exact": 0.0,
      "state_accuracy": 0.0,
      "state_all_exact": 0.0,
      "state_prefix_fraction": 0.0,
      "state_mean_correct_prefix": 0.0,
      "executor_pair_answer_consistency": 0.0,
      "executor_pair_both_correct": 0.0,
      "compiler_pair_program_consistency": 0.0,
      "compiler_pair_state_consistency": 0.0,
      "run": "pilot_expand_copy",
      "arm": "expand_copy",
      "global_step": 75,
      "stage_idx": 3
    },
    {
      "stage": "stage3_max24",
      "split": "paired_heldout_L8",
      "n": 8,
      "max_steps": 24,
      "executor_accuracy": 0.0,
      "executor_target_mass": 0.010303416289389133,
      "direct_accuracy": NaN,
      "init_accuracy": 0.0,
      "op_accuracy": 0.46875,
      "arg_accuracy": 0.0625,
      "program_exact": 0.0,
      "state_accuracy": 0.0,
      "state_all_exact": 0.0,
      "state_prefix_fraction": 0.0,
      "state_mean_correct_prefix": 0.0,
      "executor_pair_answer_consistency": 0.0,
      "executor_pair_both_correct": 0.0,
      "compiler_pair_program_consistency": 0.0,
      "compiler_pair_state_consistency": 0.0,
      "run": "pilot_expand_copy",
      "arm": "expand_copy",
      "global_step": 75,
      "stage_idx": 3
    },
    {
      "stage": "stage3_max24",
      "split": "standard_L16",
      "n": 8,
      "max_steps": 24,
      "executor_accuracy": 0.0,
      "executor_target_mass": 0.010304620955139399,
      "direct_accuracy": NaN,
      "init_accuracy": 0.0,
      "op_accuracy": 0.265625,
      "arg_accuracy": 0.0234375,
      "program_exact": 0.0,
      "state_accuracy": 0.0234375,
      "state_all_exact": 0.0,
      "state_prefix_fraction": 0.0,
      "state_mean_correct_prefix": 0.0,
      "run": "pilot_expand_copy",
      "arm": "expand_copy",
      "global_step": 75,
      "stage_idx": 3
    },
    {
      "stage": "stage3_max24",
      "split": "paraphrase_L16",
      "n": 8,
      "max_steps": 24,
      "executor_accuracy": 0.0,
      "executor_target_mass": 0.010301649570465088,
      "direct_accuracy": NaN,
      "init_accuracy": 0.0,
      "op_accuracy": 0.390625,
      "arg_accuracy": 0.0546875,
      "program_exact": 0.0,
      "state_accuracy": 0.0,
      "state_all_exact": 0.0,
      "state_prefix_fraction": 0.0,
      "state_mean_correct_prefix": 0.0,
      "run": "pilot_expand_copy",
      "arm": "expand_copy",
      "global_step": 75,
      "stage_idx": 3
    },
    {
      "stage": "stage3_max24",
      "split": "heldout_L16",
      "n": 8,
      "max_steps": 24,
      "executor_accuracy": 0.0,
      "executor_target_mass": 0.010304288938641548,
      "direct_accuracy": NaN,
      "init_accuracy": 0.0,
      "op_accuracy": 0.28125,
      "arg_accuracy": 0.03125,
      "program_exact": 0.0,
      "state_accuracy": 0.0,
      "state_all_exact": 0.0,
      "state_prefix_fraction": 0.0,
      "state_mean_correct_prefix": 0.0,
      "run": "pilot_expand_copy",
      "arm": "expand_copy",
      "global_step": 75,
      "stage_idx": 3
    },
    {
      "stage": "stage3_max24",
      "split": "paired_L16",
      "n": 8,
      "max_steps": 24,
      "executor_accuracy": 0.0,
      "executor_target_mass": 0.010298438835889101,
      "direct_accuracy": NaN,
      "init_accuracy": 0.0,
      "op_accuracy": 0.3828125,
      "arg_accuracy": 0.0625,
      "program_exact": 0.0,
      "state_accuracy": 0.0,
      "state_all_exact": 0.0,
      "state_prefix_fraction": 0.0,
      "state_mean_correct_prefix": 0.0,
      "executor_pair_answer_consistency": 0.0,
      "executor_pair_both_correct": 0.0,
      "compiler_pair_program_consistency": 0.0,
      "compiler_pair_state_consistency": 0.0,
      "run": "pilot_expand_copy",
      "arm": "expand_copy",
      "global_step": 75,
      "stage_idx": 3
    },
    {
      "stage": "stage3_max24",
      "split": "paired_heldout_L16",
      "n": 8,
      "max_steps": 24,
      "executor_accuracy": 0.0,
      "executor_target_mass": 0.010306574404239655,
      "direct_accuracy": NaN,
      "init_accuracy": 0.25,
      "op_accuracy": 0.3359375,
      "arg_accuracy": 0.015625,
      "program_exact": 0.0,
      "state_accuracy": 0.03125,
      "state_all_exact": 0.0,
      "state_prefix_fraction": 0.0,
      "state_mean_correct_prefix": 0.0,
      "executor_pair_answer_consistency": 0.0,
      "executor_pair_both_correct": 0.0,
      "compiler_pair_program_consistency": 0.0,
      "compiler_pair_state_consistency": 0.0,
      "run": "pilot_expand_copy",
      "arm": "expand_copy",
      "global_step": 75,
      "stage_idx": 3
    },
    {
      "stage": "stage3_max24",
      "split": "standard_L24",
      "n": 8,
      "max_steps": 24,
      "executor_accuracy": 0.0,
      "executor_target_mass": 0.010300551541149616,
      "direct_accuracy": NaN,
      "init_accuracy": 0.125,
      "op_accuracy": 0.3072916666666667,
      "arg_accuracy": 0.03125,
      "program_exact": 0.0,
      "state_accuracy": 0.0,
      "state_all_exact": 0.0,
      "state_prefix_fraction": 0.0,
      "state_mean_correct_prefix": 0.0,
      "run": "pilot_expand_copy",
      "arm": "expand_copy",
      "global_step": 75,
      "stage_idx": 3
    },
    {
      "stage": "stage3_max24",
      "split": "paraphrase_L24",
      "n": 8,
      "max_steps": 24,
      "executor_accuracy": 0.0,
      "executor_target_mass": 0.010397035628557205,
      "direct_accuracy": NaN,
      "init_accuracy": 0.125,
      "op_accuracy": 0.265625,
      "arg_accuracy": 0.036458333333333336,
      "program_exact": 0.0,
      "state_accuracy": 0.005208333333333333,
      "state_all_exact": 0.0,
      "state_prefix_fraction": 0.0,
      "state_mean_correct_prefix": 0.0,
      "run": "pilot_expand_copy",
      "arm": "expand_copy",
      "global_step": 75,
      "stage_idx": 3
    },
    {
      "stage": "stage3_max24",
      "split": "heldout_L24",
      "n": 8,
      "max_steps": 24,
      "executor_accuracy": 0.0,
      "executor_target_mass": 0.01030099019408226,
      "direct_accuracy": NaN,
      "init_accuracy": 0.125,
      "op_accuracy": 0.3072916666666667,
      "arg_accuracy": 0.052083333333333336,
      "program_exact": 0.0,
      "state_accuracy": 0.005208333333333333,
      "state_all_exact": 0.0,
      "state_prefix_fraction": 0.005208333333333333,
      "state_mean_correct_prefix": 0.125,
      "run": "pilot_expand_copy",
      "arm": "expand_copy",
      "global_step": 75,
      "stage_idx": 3
    },
    {
      "stage": "stage3_max24",
      "split": "paired_L24",
      "n": 8,
      "max_steps": 24,
      "executor_accuracy": 0.0,
      "executor_target_mass": 0.01029947167262435,
      "direct_accuracy": NaN,
      "init_accuracy": 0.0,
      "op_accuracy": 0.296875,
      "arg_accuracy": 0.052083333333333336,
      "program_exact": 0.0,
      "state_accuracy": 0.010416666666666666,
      "state_all_exact": 0.0,
      "state_prefix_fraction": 0.0,
      "state_mean_correct_prefix": 0.0,
      "executor_pair_answer_consistency": 0.0,
      "executor_pair_both_correct": 0.0,
      "compiler_pair_program_consistency": 0.0,
      "compiler_pair_state_consistency": 0.0,
      "run": "pilot_expand_copy",
      "arm": "expand_copy",
      "global_step": 75,
      "stage_idx": 3
    },
    {
      "stage": "stage3_max24",
      "split": "paired_heldout_L24",
      "n": 8,
      "max_steps": 24,
      "executor_accuracy": 0.0,
      "executor_target_mass": 0.01029969658702612,
      "direct_accuracy": NaN,
      "init_accuracy": 0.0,
      "op_accuracy": 0.3854166666666667,
      "arg_accuracy": 0.041666666666666664,
      "program_exact": 0.0,
      "state_accuracy": 0.005208333333333333,
      "state_all_exact": 0.0,
      "state_prefix_fraction": 0.0,
      "state_mean_correct_prefix": 0.0,
      "executor_pair_answer_consistency": 0.0,
      "executor_pair_both_correct": 0.0,
      "compiler_pair_program_consistency": 0.0,
      "compiler_pair_state_consistency": 0.0,
      "run": "pilot_expand_copy",
      "arm": "expand_copy",
      "global_step": 75,
      "stage_idx": 3
    }
  ],
  "checkpoints": [],
  "elapsed_sec": 105.49
}