{
  "run": "main_max24_no_curriculum_seed789",
  "arm": "max24_no_curriculum",
  "seed": 789,
  "args": {
    "run_name": "main_max24_no_curriculum_seed789",
    "arm": "max24_no_curriculum",
    "model_id": "Qwen/Qwen3-4B",
    "seed": 789,
    "modulus": 97,
    "stage_max_steps": "24",
    "stage_min_lengths": "1",
    "stage_train_max_lengths": "24",
    "stage_steps": "750",
    "train_examples": 512,
    "train_template_mode": "mixed",
    "eval_lengths": "8,16,24",
    "eval_examples": 64,
    "paired_eval_pairs": 32,
    "batch_size": 8,
    "grad_accum": 1,
    "eval_batch_size": 8,
    "max_length": 2048,
    "head_width": 512,
    "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": 50,
    "save_checkpoints": true,
    "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_max24",
      "max_steps": 24,
      "train_min_len": 1,
      "train_max_len": 24,
      "steps": 750
    }
  ],
  "train_log": [
    {
      "run": "main_max24_no_curriculum_seed789",
      "arm": "max24_no_curriculum",
      "seed": 789,
      "stage": "stage1_max24",
      "stage_idx": 1,
      "local_step": 1,
      "global_step": 1,
      "compiler_max_steps": 24,
      "elapsed_sec": 6.183,
      "expansion_event": "init",
      "init_loss": 4.575744152069092,
      "op_loss": 1.0922735929489136,
      "arg_loss": 4.622411251068115,
      "state_loss": 4.569647312164307,
      "state_train_accuracy": 0.02631578966975212,
      "executor_loss": 4.580330848693848,
      "loss": 47.034873962402344
    },
    {
      "run": "main_max24_no_curriculum_seed789",
      "arm": "max24_no_curriculum",
      "seed": 789,
      "stage": "stage1_max24",
      "stage_idx": 1,
      "local_step": 50,
      "global_step": 50,
      "compiler_max_steps": 24,
      "elapsed_sec": 172.372,
      "expansion_event": "init",
      "init_loss": 4.35201358795166,
      "op_loss": 1.0942409038543701,
      "arg_loss": 3.4959557056427,
      "state_loss": 4.575545787811279,
      "state_train_accuracy": 0.0,
      "executor_loss": 4.574779510498047,
      "loss": 41.636444091796875
    },
    {
      "run": "main_max24_no_curriculum_seed789",
      "arm": "max24_no_curriculum",
      "seed": 789,
      "stage": "stage1_max24",
      "stage_idx": 1,
      "local_step": 100,
      "global_step": 100,
      "compiler_max_steps": 24,
      "elapsed_sec": 343.23,
      "expansion_event": "init",
      "init_loss": 2.018043041229248,
      "op_loss": 1.0287361145019531,
      "arg_loss": 3.4977402687072754,
      "state_loss": 4.5544209480285645,
      "state_train_accuracy": 0.0,
      "executor_loss": 4.574887752532959,
      "loss": 32.2211799621582
    },
    {
      "run": "main_max24_no_curriculum_seed789",
      "arm": "max24_no_curriculum",
      "seed": 789,
      "stage": "stage1_max24",
      "stage_idx": 1,
      "local_step": 150,
      "global_step": 150,
      "compiler_max_steps": 24,
      "elapsed_sec": 515.125,
      "expansion_event": "init",
      "init_loss": 0.9972983598709106,
      "op_loss": 0.2488626092672348,
      "arg_loss": 1.3528351783752441,
      "state_loss": 4.036155700683594,
      "state_train_accuracy": 0.31496062874794006,
      "executor_loss": 4.572093963623047,
      "loss": 18.257646560668945
    },
    {
      "run": "main_max24_no_curriculum_seed789",
      "arm": "max24_no_curriculum",
      "seed": 789,
      "stage": "stage1_max24",
      "stage_idx": 1,
      "local_step": 200,
      "global_step": 200,
      "compiler_max_steps": 24,
      "elapsed_sec": 684.827,
      "expansion_event": "init",
      "init_loss": 0.11282370239496231,
      "op_loss": 9.448122727917507e-05,
      "arg_loss": 0.14024776220321655,
      "state_loss": 0.22863514721393585,
      "state_train_accuracy": 0.9898989796638489,
      "executor_loss": 0.1868043839931488,
      "loss": 1.4278199672698975
    },
    {
      "run": "main_max24_no_curriculum_seed789",
      "arm": "max24_no_curriculum",
      "seed": 789,
      "stage": "stage1_max24",
      "stage_idx": 1,
      "local_step": 250,
      "global_step": 250,
      "compiler_max_steps": 24,
      "elapsed_sec": 856.526,
      "expansion_event": "init",
      "init_loss": 0.029462363570928574,
      "op_loss": 0.06227567046880722,
      "arg_loss": 0.06495095044374466,
      "state_loss": 0.18307699263095856,
      "state_train_accuracy": 0.9837397933006287,
      "executor_loss": 1.1535502672195435,
      "loss": 1.776556134223938
    },
    {
      "run": "main_max24_no_curriculum_seed789",
      "arm": "max24_no_curriculum",
      "seed": 789,
      "stage": "stage1_max24",
      "stage_idx": 1,
      "local_step": 300,
      "global_step": 300,
      "compiler_max_steps": 24,
      "elapsed_sec": 1028.512,
      "expansion_event": "init",
      "init_loss": 0.018978511914610863,
      "op_loss": 9.370521183882374e-06,
      "arg_loss": 0.0005257942830212414,
      "state_loss": 0.01240988913923502,
      "state_train_accuracy": 1.0,
      "executor_loss": 0.024730807170271873,
      "loss": 0.1151672974228859
    },
    {
      "run": "main_max24_no_curriculum_seed789",
      "arm": "max24_no_curriculum",
      "seed": 789,
      "stage": "stage1_max24",
      "stage_idx": 1,
      "local_step": 350,
      "global_step": 350,
      "compiler_max_steps": 24,
      "elapsed_sec": 1199.809,
      "expansion_event": "init",
      "init_loss": 0.003095691092312336,
      "op_loss": 1.013947326100606e-06,
      "arg_loss": 8.33291924209334e-05,
      "state_loss": 0.003799723694100976,
      "state_train_accuracy": 1.0,
      "executor_loss": 0.004033961333334446,
      "loss": 0.02055077999830246
    },
    {
      "run": "main_max24_no_curriculum_seed789",
      "arm": "max24_no_curriculum",
      "seed": 789,
      "stage": "stage1_max24",
      "stage_idx": 1,
      "local_step": 400,
      "global_step": 400,
      "compiler_max_steps": 24,
      "elapsed_sec": 1371.374,
      "expansion_event": "init",
      "init_loss": 0.0006248244317248464,
      "op_loss": 5.834976377627754e-07,
      "arg_loss": 3.7604899262078106e-05,
      "state_loss": 0.0009397433022968471,
      "state_train_accuracy": 1.0,
      "executor_loss": 0.000987486680969596,
      "loss": 0.004577530547976494
    },
    {
      "run": "main_max24_no_curriculum_seed789",
      "arm": "max24_no_curriculum",
      "seed": 789,
      "stage": "stage1_max24",
      "stage_idx": 1,
      "local_step": 450,
      "global_step": 450,
      "compiler_max_steps": 24,
      "elapsed_sec": 1542.163,
      "expansion_event": "init",
      "init_loss": 0.003585090162232518,
      "op_loss": 3.33273135311174e-07,
      "arg_loss": 5.1017599616898224e-05,
      "state_loss": 0.0011342166690155864,
      "state_train_accuracy": 1.0,
      "executor_loss": 0.004180551040917635,
      "loss": 0.01985953189432621
    },
    {
      "run": "main_max24_no_curriculum_seed789",
      "arm": "max24_no_curriculum",
      "seed": 789,
      "stage": "stage1_max24",
      "stage_idx": 1,
      "local_step": 500,
      "global_step": 500,
      "compiler_max_steps": 24,
      "elapsed_sec": 1714.095,
      "expansion_event": "init",
      "init_loss": 0.00034788643824867904,
      "op_loss": 3.5390250729960826e-08,
      "arg_loss": 0.0002619866281747818,
      "state_loss": 0.0014906005235388875,
      "state_train_accuracy": 1.0,
      "executor_loss": 0.004538763780146837,
      "loss": 0.00846889242529869
    },
    {
      "run": "main_max24_no_curriculum_seed789",
      "arm": "max24_no_curriculum",
      "seed": 789,
      "stage": "stage1_max24",
      "stage_idx": 1,
      "local_step": 550,
      "global_step": 550,
      "compiler_max_steps": 24,
      "elapsed_sec": 1883.294,
      "expansion_event": "init",
      "init_loss": 0.00023339560721069574,
      "op_loss": 2.7509834410466283e-08,
      "arg_loss": 1.2315142157603987e-05,
      "state_loss": 0.0003142030327580869,
      "state_train_accuracy": 1.0,
      "executor_loss": 0.000373841670807451,
      "loss": 0.001670915400609374
    },
    {
      "run": "main_max24_no_curriculum_seed789",
      "arm": "max24_no_curriculum",
      "seed": 789,
      "stage": "stage1_max24",
      "stage_idx": 1,
      "local_step": 600,
      "global_step": 600,
      "compiler_max_steps": 24,
      "elapsed_sec": 2055.099,
      "expansion_event": "init",
      "init_loss": 0.00032119685783982277,
      "op_loss": 0.5168834924697876,
      "arg_loss": 1.0990674495697021,
      "state_loss": 1.7915853261947632,
      "state_train_accuracy": 0.7884615659713745,
      "executor_loss": 4.983522415161133,
      "loss": 11.689545631408691
    },
    {
      "run": "main_max24_no_curriculum_seed789",
      "arm": "max24_no_curriculum",
      "seed": 789,
      "stage": "stage1_max24",
      "stage_idx": 1,
      "local_step": 650,
      "global_step": 650,
      "compiler_max_steps": 24,
      "elapsed_sec": 2226.202,
      "expansion_event": "init",
      "init_loss": 0.0004243861767463386,
      "op_loss": 1.1582250181163545e-06,
      "arg_loss": 8.710319525562227e-05,
      "state_loss": 0.0009402563446201384,
      "state_train_accuracy": 1.0,
      "executor_loss": 0.0013953184243291616,
      "loss": 0.004382690414786339
    },
    {
      "run": "main_max24_no_curriculum_seed789",
      "arm": "max24_no_curriculum",
      "seed": 789,
      "stage": "stage1_max24",
      "stage_idx": 1,
      "local_step": 700,
      "global_step": 700,
      "compiler_max_steps": 24,
      "elapsed_sec": 2396.51,
      "expansion_event": "init",
      "init_loss": 0.0002347866538912058,
      "op_loss": 0.005980200134217739,
      "arg_loss": 0.0016504619270563126,
      "state_loss": 0.012300677597522736,
      "state_train_accuracy": 0.9999999403953552,
      "executor_loss": 0.10977621376514435,
      "loss": 0.1355980932712555
    },
    {
      "run": "main_max24_no_curriculum_seed789",
      "arm": "max24_no_curriculum",
      "seed": 789,
      "stage": "stage1_max24",
      "stage_idx": 1,
      "local_step": 750,
      "global_step": 750,
      "compiler_max_steps": 24,
      "elapsed_sec": 2566.838,
      "expansion_event": "init",
      "init_loss": 0.0003875680558849126,
      "op_loss": 0.13301794230937958,
      "arg_loss": 0.33184853196144104,
      "state_loss": 0.543584942817688,
      "state_train_accuracy": 0.8703703880310059,
      "executor_loss": 1.8098485469818115,
      "loss": 3.8153958320617676
    }
  ],
  "metrics": [
    {
      "stage": "stage1_max24",
      "split": "standard_L8",
      "n": 64,
      "max_steps": 24,
      "executor_accuracy": 1.0,
      "executor_target_mass": 0.989073283970356,
      "direct_accuracy": NaN,
      "init_accuracy": 1.0,
      "op_accuracy": 1.0,
      "arg_accuracy": 1.0,
      "program_exact": 1.0,
      "state_accuracy": 1.0,
      "state_all_exact": 1.0,
      "state_prefix_fraction": 1.0,
      "state_mean_correct_prefix": 8.0,
      "run": "main_max24_no_curriculum_seed789",
      "arm": "max24_no_curriculum",
      "seed": 789,
      "global_step": 750,
      "stage_idx": 1
    },
    {
      "stage": "stage1_max24",
      "split": "paraphrase_L8",
      "n": 64,
      "max_steps": 24,
      "executor_accuracy": 0.984375,
      "executor_target_mass": 0.9747776091098785,
      "direct_accuracy": NaN,
      "init_accuracy": 0.984375,
      "op_accuracy": 1.0,
      "arg_accuracy": 1.0,
      "program_exact": 0.984375,
      "state_accuracy": 0.984375,
      "state_all_exact": 0.984375,
      "state_prefix_fraction": 0.984375,
      "state_mean_correct_prefix": 7.875,
      "run": "main_max24_no_curriculum_seed789",
      "arm": "max24_no_curriculum",
      "seed": 789,
      "global_step": 750,
      "stage_idx": 1
    },
    {
      "stage": "stage1_max24",
      "split": "heldout_L8",
      "n": 64,
      "max_steps": 24,
      "executor_accuracy": 0.921875,
      "executor_target_mass": 0.9008565619587898,
      "direct_accuracy": NaN,
      "init_accuracy": 0.984375,
      "op_accuracy": 0.962890625,
      "arg_accuracy": 0.95703125,
      "program_exact": 0.921875,
      "state_accuracy": 0.947265625,
      "state_all_exact": 0.921875,
      "state_prefix_fraction": 0.947265625,
      "state_mean_correct_prefix": 7.578125,
      "run": "main_max24_no_curriculum_seed789",
      "arm": "max24_no_curriculum",
      "seed": 789,
      "global_step": 750,
      "stage_idx": 1
    },
    {
      "stage": "stage1_max24",
      "split": "paired_L8",
      "n": 64,
      "max_steps": 24,
      "executor_accuracy": 1.0,
      "executor_target_mass": 0.9892675653100014,
      "direct_accuracy": NaN,
      "init_accuracy": 1.0,
      "op_accuracy": 1.0,
      "arg_accuracy": 1.0,
      "program_exact": 1.0,
      "state_accuracy": 1.0,
      "state_all_exact": 1.0,
      "state_prefix_fraction": 1.0,
      "state_mean_correct_prefix": 8.0,
      "executor_pair_answer_consistency": 1.0,
      "executor_pair_both_correct": 1.0,
      "compiler_pair_program_consistency": 1.0,
      "compiler_pair_state_consistency": 1.0,
      "run": "main_max24_no_curriculum_seed789",
      "arm": "max24_no_curriculum",
      "seed": 789,
      "global_step": 750,
      "stage_idx": 1
    },
    {
      "stage": "stage1_max24",
      "split": "paired_heldout_L8",
      "n": 64,
      "max_steps": 24,
      "executor_accuracy": 0.90625,
      "executor_target_mass": 0.8947992660105228,
      "direct_accuracy": NaN,
      "init_accuracy": 0.9375,
      "op_accuracy": 0.984375,
      "arg_accuracy": 0.978515625,
      "program_exact": 0.90625,
      "state_accuracy": 0.9140625,
      "state_all_exact": 0.90625,
      "state_prefix_fraction": 0.9140625,
      "state_mean_correct_prefix": 7.3125,
      "executor_pair_answer_consistency": 0.90625,
      "executor_pair_both_correct": 0.875,
      "compiler_pair_program_consistency": 0.90625,
      "compiler_pair_state_consistency": 0.90625,
      "run": "main_max24_no_curriculum_seed789",
      "arm": "max24_no_curriculum",
      "seed": 789,
      "global_step": 750,
      "stage_idx": 1
    },
    {
      "stage": "stage1_max24",
      "split": "standard_L16",
      "n": 64,
      "max_steps": 24,
      "executor_accuracy": 0.96875,
      "executor_target_mass": 0.9116620793938637,
      "direct_accuracy": NaN,
      "init_accuracy": 0.96875,
      "op_accuracy": 1.0,
      "arg_accuracy": 1.0,
      "program_exact": 0.96875,
      "state_accuracy": 0.96875,
      "state_all_exact": 0.96875,
      "state_prefix_fraction": 0.96875,
      "state_mean_correct_prefix": 15.5,
      "run": "main_max24_no_curriculum_seed789",
      "arm": "max24_no_curriculum",
      "seed": 789,
      "global_step": 750,
      "stage_idx": 1
    },
    {
      "stage": "stage1_max24",
      "split": "paraphrase_L16",
      "n": 64,
      "max_steps": 24,
      "executor_accuracy": 0.984375,
      "executor_target_mass": 0.930874228477478,
      "direct_accuracy": NaN,
      "init_accuracy": 0.984375,
      "op_accuracy": 1.0,
      "arg_accuracy": 1.0,
      "program_exact": 0.984375,
      "state_accuracy": 0.984375,
      "state_all_exact": 0.984375,
      "state_prefix_fraction": 0.984375,
      "state_mean_correct_prefix": 15.75,
      "run": "main_max24_no_curriculum_seed789",
      "arm": "max24_no_curriculum",
      "seed": 789,
      "global_step": 750,
      "stage_idx": 1
    },
    {
      "stage": "stage1_max24",
      "split": "heldout_L16",
      "n": 64,
      "max_steps": 24,
      "executor_accuracy": 0.90625,
      "executor_target_mass": 0.8499426320195198,
      "direct_accuracy": NaN,
      "init_accuracy": 0.984375,
      "op_accuracy": 0.955078125,
      "arg_accuracy": 0.94921875,
      "program_exact": 0.90625,
      "state_accuracy": 0.919921875,
      "state_all_exact": 0.90625,
      "state_prefix_fraction": 0.9189453125,
      "state_mean_correct_prefix": 14.703125,
      "run": "main_max24_no_curriculum_seed789",
      "arm": "max24_no_curriculum",
      "seed": 789,
      "global_step": 750,
      "stage_idx": 1
    },
    {
      "stage": "stage1_max24",
      "split": "paired_L16",
      "n": 64,
      "max_steps": 24,
      "executor_accuracy": 0.984375,
      "executor_target_mass": 0.9225541427731514,
      "direct_accuracy": NaN,
      "init_accuracy": 1.0,
      "op_accuracy": 1.0,
      "arg_accuracy": 0.9990234375,
      "program_exact": 0.984375,
      "state_accuracy": 0.998046875,
      "state_all_exact": 0.984375,
      "state_prefix_fraction": 0.998046875,
      "state_mean_correct_prefix": 15.96875,
      "executor_pair_answer_consistency": 0.96875,
      "executor_pair_both_correct": 0.96875,
      "compiler_pair_program_consistency": 0.96875,
      "compiler_pair_state_consistency": 0.96875,
      "run": "main_max24_no_curriculum_seed789",
      "arm": "max24_no_curriculum",
      "seed": 789,
      "global_step": 750,
      "stage_idx": 1
    },
    {
      "stage": "stage1_max24",
      "split": "paired_heldout_L16",
      "n": 64,
      "max_steps": 24,
      "executor_accuracy": 0.953125,
      "executor_target_mass": 0.888838030397892,
      "direct_accuracy": NaN,
      "init_accuracy": 1.0,
      "op_accuracy": 0.9716796875,
      "arg_accuracy": 0.962890625,
      "program_exact": 0.953125,
      "state_accuracy": 0.9619140625,
      "state_all_exact": 0.953125,
      "state_prefix_fraction": 0.9619140625,
      "state_mean_correct_prefix": 15.390625,
      "executor_pair_answer_consistency": 0.90625,
      "executor_pair_both_correct": 0.90625,
      "compiler_pair_program_consistency": 0.90625,
      "compiler_pair_state_consistency": 0.90625,
      "run": "main_max24_no_curriculum_seed789",
      "arm": "max24_no_curriculum",
      "seed": 789,
      "global_step": 750,
      "stage_idx": 1
    },
    {
      "stage": "stage1_max24",
      "split": "standard_L24",
      "n": 64,
      "max_steps": 24,
      "executor_accuracy": 0.015625,
      "executor_target_mass": 0.010285425116308033,
      "direct_accuracy": NaN,
      "init_accuracy": 1.0,
      "op_accuracy": 0.8274739583333334,
      "arg_accuracy": 0.7408854166666666,
      "program_exact": 0.0,
      "state_accuracy": 0.6901041666666666,
      "state_all_exact": 0.0,
      "state_prefix_fraction": 0.6875,
      "state_mean_correct_prefix": 16.5,
      "run": "main_max24_no_curriculum_seed789",
      "arm": "max24_no_curriculum",
      "seed": 789,
      "global_step": 750,
      "stage_idx": 1
    },
    {
      "stage": "stage1_max24",
      "split": "paraphrase_L24",
      "n": 64,
      "max_steps": 24,
      "executor_accuracy": 0.0,
      "executor_target_mass": 0.010369550669565797,
      "direct_accuracy": NaN,
      "init_accuracy": 1.0,
      "op_accuracy": 0.880859375,
      "arg_accuracy": 0.8372395833333334,
      "program_exact": 0.0,
      "state_accuracy": 0.7786458333333334,
      "state_all_exact": 0.0,
      "state_prefix_fraction": 0.775390625,
      "state_mean_correct_prefix": 18.609375,
      "run": "main_max24_no_curriculum_seed789",
      "arm": "max24_no_curriculum",
      "seed": 789,
      "global_step": 750,
      "stage_idx": 1
    },
    {
      "stage": "stage1_max24",
      "split": "heldout_L24",
      "n": 64,
      "max_steps": 24,
      "executor_accuracy": 0.0,
      "executor_target_mass": 0.010319428285583854,
      "direct_accuracy": NaN,
      "init_accuracy": 1.0,
      "op_accuracy": 0.83984375,
      "arg_accuracy": 0.7623697916666666,
      "program_exact": 0.0,
      "state_accuracy": 0.7109375,
      "state_all_exact": 0.0,
      "state_prefix_fraction": 0.7076822916666666,
      "state_mean_correct_prefix": 16.984375,
      "run": "main_max24_no_curriculum_seed789",
      "arm": "max24_no_curriculum",
      "seed": 789,
      "global_step": 750,
      "stage_idx": 1
    },
    {
      "stage": "stage1_max24",
      "split": "paired_L24",
      "n": 64,
      "max_steps": 24,
      "executor_accuracy": 0.0,
      "executor_target_mass": 0.010348917334340513,
      "direct_accuracy": NaN,
      "init_accuracy": 1.0,
      "op_accuracy": 0.8717447916666666,
      "arg_accuracy": 0.783203125,
      "program_exact": 0.0,
      "state_accuracy": 0.7350260416666666,
      "state_all_exact": 0.0,
      "state_prefix_fraction": 0.7330729166666666,
      "state_mean_correct_prefix": 17.59375,
      "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": "main_max24_no_curriculum_seed789",
      "arm": "max24_no_curriculum",
      "seed": 789,
      "global_step": 750,
      "stage_idx": 1
    },
    {
      "stage": "stage1_max24",
      "split": "paired_heldout_L24",
      "n": 64,
      "max_steps": 24,
      "executor_accuracy": 0.0,
      "executor_target_mass": 0.010317265056073666,
      "direct_accuracy": NaN,
      "init_accuracy": 0.96875,
      "op_accuracy": 0.8190104166666666,
      "arg_accuracy": 0.740234375,
      "program_exact": 0.0,
      "state_accuracy": 0.6634114583333334,
      "state_all_exact": 0.0,
      "state_prefix_fraction": 0.66015625,
      "state_mean_correct_prefix": 15.84375,
      "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": "main_max24_no_curriculum_seed789",
      "arm": "max24_no_curriculum",
      "seed": 789,
      "global_step": 750,
      "stage_idx": 1
    }
  ],
  "checkpoints": [
    {
      "run": "main_max24_no_curriculum_seed789",
      "seed": 789,
      "tag": "stage1_max24_step00750",
      "stage": "stage1_max24",
      "global_step": 750,
      "compiler_max_steps": 24,
      "checkpoint_dir": "/workspace/large_artifacts/qwen_compiler_multiseed_reattribution/checkpoints/main_max24_no_curriculum_seed789/stage1_max24_step00750",
      "model_checkpoint": "/workspace/large_artifacts/qwen_compiler_multiseed_reattribution/checkpoints/main_max24_no_curriculum_seed789/stage1_max24_step00750/adapter",
      "heads_checkpoint": "/workspace/large_artifacts/qwen_compiler_multiseed_reattribution/checkpoints/main_max24_no_curriculum_seed789/stage1_max24_step00750/heads.pt",
      "selection_split": "paired_L24",
      "selection_metric": "executor_pair_both_correct",
      "selection_value": 0.0,
      "executor_accuracy": 0.0,
      "program_exact": 0.0,
      "state_prefix_fraction": 0.7330729166666666
    }
  ],
  "elapsed_sec": 2694.57
}