{
  "run": "smoke_random_new_slots",
  "arm": "smoke_random_new_slots",
  "args": {
    "run_name": "smoke_random_new_slots",
    "arm": "smoke_random_new_slots",
    "model_id": "Qwen/Qwen3-4B",
    "seed": 123,
    "modulus": 97,
    "stage_max_steps": "8,16,24",
    "stage_min_lengths": "1,1,1",
    "stage_train_max_lengths": "4,8,12",
    "stage_steps": "1,1,1",
    "train_examples": 4,
    "train_template_mode": "mixed",
    "eval_lengths": "8,16,24",
    "eval_examples": 2,
    "paired_eval_pairs": 1,
    "batch_size": 1,
    "grad_accum": 1,
    "eval_batch_size": 1,
    "max_length": 2048,
    "head_width": 128,
    "compiler_layers": 1,
    "compiler_heads": 4,
    "compiler_dropout": 0.05,
    "expansion_noise": 0.005,
    "expansion_mode": "random_new",
    "trace_loss_weight": 1.0,
    "executor_loss_weight": 1.0,
    "state_loss_weight": 0.25,
    "init_trace_loss_weight": 1.0,
    "op_trace_loss_weight": 1.0,
    "arg_trace_loss_weight": 1.0,
    "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": 16,
    "lora_alpha": 32,
    "lora_dropout": 0.05,
    "lora_target_modules": "all-linear",
    "log_interval": 1,
    "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": 16,
    "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": 4,
      "steps": 1
    },
    {
      "name": "stage2_max16",
      "max_steps": 16,
      "train_min_len": 1,
      "train_max_len": 8,
      "steps": 1
    },
    {
      "name": "stage3_max24",
      "max_steps": 24,
      "train_min_len": 1,
      "train_max_len": 12,
      "steps": 1
    }
  ],
  "train_log": [
    {
      "run": "smoke_random_new_slots",
      "arm": "smoke_random_new_slots",
      "stage": "stage1_max8",
      "stage_idx": 1,
      "local_step": 1,
      "global_step": 1,
      "compiler_max_steps": 8,
      "elapsed_sec": 1.066,
      "expansion_event": "init",
      "init_loss": 4.608971118927002,
      "op_loss": 1.0004818439483643,
      "arg_loss": 4.786462783813477,
      "state_loss": 4.580732345581055,
      "state_train_accuracy": 0.0,
      "executor_loss": 4.580732345581055,
      "loss": 16.1218318939209
    },
    {
      "run": "smoke_random_new_slots",
      "arm": "smoke_random_new_slots",
      "stage": "stage2_max16",
      "stage_idx": 2,
      "local_step": 1,
      "global_step": 2,
      "compiler_max_steps": 16,
      "elapsed_sec": 2.785,
      "expansion_event": "expand_8_to_16_random_new",
      "init_loss": 4.500669002532959,
      "op_loss": 1.2040961980819702,
      "arg_loss": 4.596749305725098,
      "state_loss": 4.578803062438965,
      "state_train_accuracy": 0.0,
      "executor_loss": 4.579156398773193,
      "loss": 16.025371551513672
    },
    {
      "run": "smoke_random_new_slots",
      "arm": "smoke_random_new_slots",
      "stage": "stage3_max24",
      "stage_idx": 3,
      "local_step": 1,
      "global_step": 3,
      "compiler_max_steps": 24,
      "elapsed_sec": 5.854,
      "expansion_event": "expand_16_to_24_random_new",
      "init_loss": 4.571833610534668,
      "op_loss": 1.1309095621109009,
      "arg_loss": 4.655930519104004,
      "state_loss": 4.580975532531738,
      "state_train_accuracy": 0.0,
      "executor_loss": 4.580688953399658,
      "loss": 16.084606170654297
    }
  ],
  "metrics": [
    {
      "stage": "stage1_max8",
      "split": "standard_L8",
      "n": 2,
      "max_steps": 8,
      "executor_accuracy": 0.0,
      "executor_target_mass": 0.010260722134262323,
      "direct_accuracy": NaN,
      "init_accuracy": 0.0,
      "op_accuracy": 0.5,
      "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,
      "run": "smoke_random_new_slots",
      "arm": "smoke_random_new_slots",
      "global_step": 1,
      "stage_idx": 1
    },
    {
      "stage": "stage1_max8",
      "split": "paraphrase_L8",
      "n": 2,
      "max_steps": 8,
      "executor_accuracy": 0.0,
      "executor_target_mass": 0.010262320283800364,
      "direct_accuracy": NaN,
      "init_accuracy": 0.0,
      "op_accuracy": 0.375,
      "arg_accuracy": 0.0,
      "program_exact": 0.0,
      "state_accuracy": 0.0625,
      "state_all_exact": 0.0,
      "state_prefix_fraction": 0.0,
      "state_mean_correct_prefix": 0.0,
      "run": "smoke_random_new_slots",
      "arm": "smoke_random_new_slots",
      "global_step": 1,
      "stage_idx": 1
    },
    {
      "stage": "stage1_max8",
      "split": "heldout_L8",
      "n": 2,
      "max_steps": 8,
      "executor_accuracy": 0.0,
      "executor_target_mass": 0.01026733499020338,
      "direct_accuracy": NaN,
      "init_accuracy": 0.0,
      "op_accuracy": 0.1875,
      "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,
      "run": "smoke_random_new_slots",
      "arm": "smoke_random_new_slots",
      "global_step": 1,
      "stage_idx": 1
    },
    {
      "stage": "stage1_max8",
      "split": "paired_L8",
      "n": 2,
      "max_steps": 8,
      "executor_accuracy": 0.0,
      "executor_target_mass": 0.010258032940328121,
      "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": 1.0,
      "executor_pair_both_correct": 0.0,
      "compiler_pair_program_consistency": 1.0,
      "compiler_pair_state_consistency": 1.0,
      "run": "smoke_random_new_slots",
      "arm": "smoke_random_new_slots",
      "global_step": 1,
      "stage_idx": 1
    },
    {
      "stage": "stage1_max8",
      "split": "paired_heldout_L8",
      "n": 2,
      "max_steps": 8,
      "executor_accuracy": 0.0,
      "executor_target_mass": 0.010270128026604652,
      "direct_accuracy": NaN,
      "init_accuracy": 0.0,
      "op_accuracy": 0.375,
      "arg_accuracy": 0.0,
      "program_exact": 0.0,
      "state_accuracy": 0.0625,
      "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": "smoke_random_new_slots",
      "arm": "smoke_random_new_slots",
      "global_step": 1,
      "stage_idx": 1
    },
    {
      "stage": "stage2_max16",
      "split": "standard_L8",
      "n": 2,
      "max_steps": 16,
      "executor_accuracy": 0.0,
      "executor_target_mass": 0.010250100865960121,
      "direct_accuracy": NaN,
      "init_accuracy": 0.0,
      "op_accuracy": 0.25,
      "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": "smoke_random_new_slots",
      "arm": "smoke_random_new_slots",
      "global_step": 2,
      "stage_idx": 2
    },
    {
      "stage": "stage2_max16",
      "split": "paraphrase_L8",
      "n": 2,
      "max_steps": 16,
      "executor_accuracy": 0.0,
      "executor_target_mass": 0.01024750480428338,
      "direct_accuracy": NaN,
      "init_accuracy": 0.0,
      "op_accuracy": 0.125,
      "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,
      "run": "smoke_random_new_slots",
      "arm": "smoke_random_new_slots",
      "global_step": 2,
      "stage_idx": 2
    },
    {
      "stage": "stage2_max16",
      "split": "heldout_L8",
      "n": 2,
      "max_steps": 16,
      "executor_accuracy": 0.0,
      "executor_target_mass": 0.010236272122710943,
      "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,
      "run": "smoke_random_new_slots",
      "arm": "smoke_random_new_slots",
      "global_step": 2,
      "stage_idx": 2
    },
    {
      "stage": "stage2_max16",
      "split": "paired_L8",
      "n": 2,
      "max_steps": 16,
      "executor_accuracy": 0.0,
      "executor_target_mass": 0.010246952064335346,
      "direct_accuracy": NaN,
      "init_accuracy": 0.0,
      "op_accuracy": 0.5625,
      "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": "smoke_random_new_slots",
      "arm": "smoke_random_new_slots",
      "global_step": 2,
      "stage_idx": 2
    },
    {
      "stage": "stage2_max16",
      "split": "paired_heldout_L8",
      "n": 2,
      "max_steps": 16,
      "executor_accuracy": 0.0,
      "executor_target_mass": 0.01024617999792099,
      "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": 0.0,
      "executor_pair_both_correct": 0.0,
      "compiler_pair_program_consistency": 0.0,
      "compiler_pair_state_consistency": 0.0,
      "run": "smoke_random_new_slots",
      "arm": "smoke_random_new_slots",
      "global_step": 2,
      "stage_idx": 2
    },
    {
      "stage": "stage2_max16",
      "split": "standard_L16",
      "n": 2,
      "max_steps": 16,
      "executor_accuracy": 0.0,
      "executor_target_mass": 0.01023790379986167,
      "direct_accuracy": NaN,
      "init_accuracy": 0.0,
      "op_accuracy": 0.3125,
      "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,
      "run": "smoke_random_new_slots",
      "arm": "smoke_random_new_slots",
      "global_step": 2,
      "stage_idx": 2
    },
    {
      "stage": "stage2_max16",
      "split": "paraphrase_L16",
      "n": 2,
      "max_steps": 16,
      "executor_accuracy": 0.0,
      "executor_target_mass": 0.010232329368591309,
      "direct_accuracy": NaN,
      "init_accuracy": 0.0,
      "op_accuracy": 0.40625,
      "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,
      "run": "smoke_random_new_slots",
      "arm": "smoke_random_new_slots",
      "global_step": 2,
      "stage_idx": 2
    },
    {
      "stage": "stage2_max16",
      "split": "heldout_L16",
      "n": 2,
      "max_steps": 16,
      "executor_accuracy": 0.0,
      "executor_target_mass": 0.010244045406579971,
      "direct_accuracy": NaN,
      "init_accuracy": 0.0,
      "op_accuracy": 0.40625,
      "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,
      "run": "smoke_random_new_slots",
      "arm": "smoke_random_new_slots",
      "global_step": 2,
      "stage_idx": 2
    },
    {
      "stage": "stage2_max16",
      "split": "paired_L16",
      "n": 2,
      "max_steps": 16,
      "executor_accuracy": 0.0,
      "executor_target_mass": 0.01023393776267767,
      "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": 0.0,
      "executor_pair_both_correct": 0.0,
      "compiler_pair_program_consistency": 0.0,
      "compiler_pair_state_consistency": 0.0,
      "run": "smoke_random_new_slots",
      "arm": "smoke_random_new_slots",
      "global_step": 2,
      "stage_idx": 2
    },
    {
      "stage": "stage2_max16",
      "split": "paired_heldout_L16",
      "n": 2,
      "max_steps": 16,
      "executor_accuracy": 0.0,
      "executor_target_mass": 0.010241682175546885,
      "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": 0.0,
      "executor_pair_both_correct": 0.0,
      "compiler_pair_program_consistency": 0.0,
      "compiler_pair_state_consistency": 0.0,
      "run": "smoke_random_new_slots",
      "arm": "smoke_random_new_slots",
      "global_step": 2,
      "stage_idx": 2
    },
    {
      "stage": "stage3_max24",
      "split": "standard_L8",
      "n": 2,
      "max_steps": 24,
      "executor_accuracy": 0.0,
      "executor_target_mass": 0.010236228816211224,
      "direct_accuracy": NaN,
      "init_accuracy": 0.0,
      "op_accuracy": 0.3125,
      "arg_accuracy": 0.1875,
      "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": "smoke_random_new_slots",
      "arm": "smoke_random_new_slots",
      "global_step": 3,
      "stage_idx": 3
    },
    {
      "stage": "stage3_max24",
      "split": "paraphrase_L8",
      "n": 2,
      "max_steps": 24,
      "executor_accuracy": 0.0,
      "executor_target_mass": 0.010241515934467316,
      "direct_accuracy": NaN,
      "init_accuracy": 0.0,
      "op_accuracy": 0.375,
      "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,
      "run": "smoke_random_new_slots",
      "arm": "smoke_random_new_slots",
      "global_step": 3,
      "stage_idx": 3
    },
    {
      "stage": "stage3_max24",
      "split": "heldout_L8",
      "n": 2,
      "max_steps": 24,
      "executor_accuracy": 0.0,
      "executor_target_mass": 0.010242386721074581,
      "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": "smoke_random_new_slots",
      "arm": "smoke_random_new_slots",
      "global_step": 3,
      "stage_idx": 3
    },
    {
      "stage": "stage3_max24",
      "split": "paired_L8",
      "n": 2,
      "max_steps": 24,
      "executor_accuracy": 0.0,
      "executor_target_mass": 0.010245692916214466,
      "direct_accuracy": NaN,
      "init_accuracy": 0.0,
      "op_accuracy": 0.375,
      "arg_accuracy": 0.125,
      "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": "smoke_random_new_slots",
      "arm": "smoke_random_new_slots",
      "global_step": 3,
      "stage_idx": 3
    },
    {
      "stage": "stage3_max24",
      "split": "paired_heldout_L8",
      "n": 2,
      "max_steps": 24,
      "executor_accuracy": 0.0,
      "executor_target_mass": 0.010235333815217018,
      "direct_accuracy": NaN,
      "init_accuracy": 0.0,
      "op_accuracy": 0.75,
      "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": "smoke_random_new_slots",
      "arm": "smoke_random_new_slots",
      "global_step": 3,
      "stage_idx": 3
    },
    {
      "stage": "stage3_max24",
      "split": "standard_L16",
      "n": 2,
      "max_steps": 24,
      "executor_accuracy": 0.0,
      "executor_target_mass": 0.010237665381282568,
      "direct_accuracy": NaN,
      "init_accuracy": 0.0,
      "op_accuracy": 0.21875,
      "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,
      "run": "smoke_random_new_slots",
      "arm": "smoke_random_new_slots",
      "global_step": 3,
      "stage_idx": 3
    },
    {
      "stage": "stage3_max24",
      "split": "paraphrase_L16",
      "n": 2,
      "max_steps": 24,
      "executor_accuracy": 0.0,
      "executor_target_mass": 0.010246165562421083,
      "direct_accuracy": NaN,
      "init_accuracy": 0.0,
      "op_accuracy": 0.40625,
      "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": "smoke_random_new_slots",
      "arm": "smoke_random_new_slots",
      "global_step": 3,
      "stage_idx": 3
    },
    {
      "stage": "stage3_max24",
      "split": "heldout_L16",
      "n": 2,
      "max_steps": 24,
      "executor_accuracy": 0.0,
      "executor_target_mass": 0.010241229087114334,
      "direct_accuracy": NaN,
      "init_accuracy": 0.0,
      "op_accuracy": 0.28125,
      "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,
      "run": "smoke_random_new_slots",
      "arm": "smoke_random_new_slots",
      "global_step": 3,
      "stage_idx": 3
    },
    {
      "stage": "stage3_max24",
      "split": "paired_L16",
      "n": 2,
      "max_steps": 24,
      "executor_accuracy": 0.0,
      "executor_target_mass": 0.010245639365166426,
      "direct_accuracy": NaN,
      "init_accuracy": 0.0,
      "op_accuracy": 0.5,
      "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": "smoke_random_new_slots",
      "arm": "smoke_random_new_slots",
      "global_step": 3,
      "stage_idx": 3
    },
    {
      "stage": "stage3_max24",
      "split": "paired_heldout_L16",
      "n": 2,
      "max_steps": 24,
      "executor_accuracy": 0.0,
      "executor_target_mass": 0.010246395133435726,
      "direct_accuracy": NaN,
      "init_accuracy": 0.0,
      "op_accuracy": 0.375,
      "arg_accuracy": 0.125,
      "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": "smoke_random_new_slots",
      "arm": "smoke_random_new_slots",
      "global_step": 3,
      "stage_idx": 3
    },
    {
      "stage": "stage3_max24",
      "split": "standard_L24",
      "n": 2,
      "max_steps": 24,
      "executor_accuracy": 0.0,
      "executor_target_mass": 0.010246035642921925,
      "direct_accuracy": NaN,
      "init_accuracy": 0.0,
      "op_accuracy": 0.3125,
      "arg_accuracy": 0.020833333333333332,
      "program_exact": 0.0,
      "state_accuracy": 0.020833333333333332,
      "state_all_exact": 0.0,
      "state_prefix_fraction": 0.020833333333333332,
      "state_mean_correct_prefix": 0.5,
      "run": "smoke_random_new_slots",
      "arm": "smoke_random_new_slots",
      "global_step": 3,
      "stage_idx": 3
    },
    {
      "stage": "stage3_max24",
      "split": "paraphrase_L24",
      "n": 2,
      "max_steps": 24,
      "executor_accuracy": 0.0,
      "executor_target_mass": 0.010233789216727018,
      "direct_accuracy": NaN,
      "init_accuracy": 0.0,
      "op_accuracy": 0.22916666666666666,
      "arg_accuracy": 0.041666666666666664,
      "program_exact": 0.0,
      "state_accuracy": 0.041666666666666664,
      "state_all_exact": 0.0,
      "state_prefix_fraction": 0.0,
      "state_mean_correct_prefix": 0.0,
      "run": "smoke_random_new_slots",
      "arm": "smoke_random_new_slots",
      "global_step": 3,
      "stage_idx": 3
    },
    {
      "stage": "stage3_max24",
      "split": "heldout_L24",
      "n": 2,
      "max_steps": 24,
      "executor_accuracy": 0.0,
      "executor_target_mass": 0.010238362941890955,
      "direct_accuracy": NaN,
      "init_accuracy": 0.0,
      "op_accuracy": 0.2916666666666667,
      "arg_accuracy": 0.020833333333333332,
      "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": "smoke_random_new_slots",
      "arm": "smoke_random_new_slots",
      "global_step": 3,
      "stage_idx": 3
    },
    {
      "stage": "stage3_max24",
      "split": "paired_L24",
      "n": 2,
      "max_steps": 24,
      "executor_accuracy": 0.0,
      "executor_target_mass": 0.01023798156529665,
      "direct_accuracy": NaN,
      "init_accuracy": 0.0,
      "op_accuracy": 0.3333333333333333,
      "arg_accuracy": 0.041666666666666664,
      "program_exact": 0.0,
      "state_accuracy": 0.020833333333333332,
      "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": "smoke_random_new_slots",
      "arm": "smoke_random_new_slots",
      "global_step": 3,
      "stage_idx": 3
    },
    {
      "stage": "stage3_max24",
      "split": "paired_heldout_L24",
      "n": 2,
      "max_steps": 24,
      "executor_accuracy": 0.0,
      "executor_target_mass": 0.010251238010823727,
      "direct_accuracy": NaN,
      "init_accuracy": 0.0,
      "op_accuracy": 0.375,
      "arg_accuracy": 0.041666666666666664,
      "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": "smoke_random_new_slots",
      "arm": "smoke_random_new_slots",
      "global_step": 3,
      "stage_idx": 3
    }
  ],
  "checkpoints": [],
  "elapsed_sec": 10.157
}