{
  "run": "smoke_qwen_structural_expansion",
  "args": {
    "run_name": "smoke_qwen_structural_expansion",
    "model_id": "Qwen/Qwen3-4B",
    "seed": 123,
    "modulus": 97,
    "stage_max_steps": "8,16,24",
    "stage_min_lengths": "1,8,16",
    "stage_train_max_lengths": "8,16,24",
    "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,
    "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": 8,
      "steps": 1
    },
    {
      "name": "stage2_max16",
      "max_steps": 16,
      "train_min_len": 8,
      "train_max_len": 16,
      "steps": 1
    },
    {
      "name": "stage3_max24",
      "max_steps": 24,
      "train_min_len": 16,
      "train_max_len": 24,
      "steps": 1
    }
  ],
  "train_log": [
    {
      "run": "smoke_qwen_structural_expansion",
      "stage": "stage1_max8",
      "stage_idx": 1,
      "local_step": 1,
      "global_step": 1,
      "compiler_max_steps": 8,
      "elapsed_sec": 1.083,
      "expansion_event": "init",
      "init_loss": 4.627433776855469,
      "op_loss": 1.0218236446380615,
      "arg_loss": 4.819574356079102,
      "state_loss": 4.579990386962891,
      "state_train_accuracy": 0.0,
      "executor_loss": 4.5796217918396,
      "loss": 16.193450927734375
    },
    {
      "run": "smoke_qwen_structural_expansion",
      "stage": "stage2_max16",
      "stage_idx": 2,
      "local_step": 1,
      "global_step": 2,
      "compiler_max_steps": 16,
      "elapsed_sec": 2.366,
      "expansion_event": "expand_8_to_16",
      "init_loss": 4.443595886230469,
      "op_loss": 1.0334206819534302,
      "arg_loss": 4.619872093200684,
      "state_loss": 4.5811357498168945,
      "state_train_accuracy": 0.0,
      "executor_loss": 4.581022262573242,
      "loss": 15.82319450378418
    },
    {
      "run": "smoke_qwen_structural_expansion",
      "stage": "stage3_max24",
      "stage_idx": 3,
      "local_step": 1,
      "global_step": 3,
      "compiler_max_steps": 24,
      "elapsed_sec": 4.421,
      "expansion_event": "expand_16_to_24",
      "init_loss": 4.471540451049805,
      "op_loss": 1.1556838750839233,
      "arg_loss": 4.533812999725342,
      "state_loss": 4.580985069274902,
      "state_train_accuracy": 0.0,
      "executor_loss": 4.581622123718262,
      "loss": 15.887906074523926
    }
  ],
  "metrics": [
    {
      "stage": "stage1_max8",
      "split": "standard_L8",
      "n": 2,
      "max_steps": 8,
      "executor_accuracy": 0.0,
      "executor_target_mass": 0.01023809053003788,
      "direct_accuracy": NaN,
      "init_accuracy": 0.0,
      "op_accuracy": 0.625,
      "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_qwen_structural_expansion",
      "global_step": 1,
      "stage_idx": 1
    },
    {
      "stage": "stage1_max8",
      "split": "paraphrase_L8",
      "n": 2,
      "max_steps": 8,
      "executor_accuracy": 0.0,
      "executor_target_mass": 0.010238724760711193,
      "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_qwen_structural_expansion",
      "global_step": 1,
      "stage_idx": 1
    },
    {
      "stage": "stage1_max8",
      "split": "paired_L8",
      "n": 2,
      "max_steps": 8,
      "executor_accuracy": 0.0,
      "executor_target_mass": 0.010232257191091776,
      "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_qwen_structural_expansion",
      "global_step": 1,
      "stage_idx": 1
    },
    {
      "stage": "stage2_max16",
      "split": "standard_L8",
      "n": 2,
      "max_steps": 16,
      "executor_accuracy": 0.0,
      "executor_target_mass": 0.010247760452330112,
      "direct_accuracy": NaN,
      "init_accuracy": 0.0,
      "op_accuracy": 0.625,
      "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_qwen_structural_expansion",
      "global_step": 2,
      "stage_idx": 2
    },
    {
      "stage": "stage2_max16",
      "split": "paraphrase_L8",
      "n": 2,
      "max_steps": 16,
      "executor_accuracy": 0.0,
      "executor_target_mass": 0.010244766715914011,
      "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_qwen_structural_expansion",
      "global_step": 2,
      "stage_idx": 2
    },
    {
      "stage": "stage2_max16",
      "split": "paired_L8",
      "n": 2,
      "max_steps": 16,
      "executor_accuracy": 0.0,
      "executor_target_mass": 0.010240803472697735,
      "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,
      "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_qwen_structural_expansion",
      "global_step": 2,
      "stage_idx": 2
    },
    {
      "stage": "stage2_max16",
      "split": "standard_L16",
      "n": 2,
      "max_steps": 16,
      "executor_accuracy": 0.0,
      "executor_target_mass": 0.010238269809633493,
      "direct_accuracy": NaN,
      "init_accuracy": 0.0,
      "op_accuracy": 0.25,
      "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": "smoke_qwen_structural_expansion",
      "global_step": 2,
      "stage_idx": 2
    },
    {
      "stage": "stage2_max16",
      "split": "paraphrase_L16",
      "n": 2,
      "max_steps": 16,
      "executor_accuracy": 0.0,
      "executor_target_mass": 0.010245541576296091,
      "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_qwen_structural_expansion",
      "global_step": 2,
      "stage_idx": 2
    },
    {
      "stage": "stage2_max16",
      "split": "paired_L16",
      "n": 2,
      "max_steps": 16,
      "executor_accuracy": 0.0,
      "executor_target_mass": 0.01024319976568222,
      "direct_accuracy": NaN,
      "init_accuracy": 0.0,
      "op_accuracy": 0.25,
      "arg_accuracy": 0.0625,
      "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_qwen_structural_expansion",
      "global_step": 2,
      "stage_idx": 2
    },
    {
      "stage": "stage3_max24",
      "split": "standard_L8",
      "n": 2,
      "max_steps": 24,
      "executor_accuracy": 0.0,
      "executor_target_mass": 0.010255424305796623,
      "direct_accuracy": NaN,
      "init_accuracy": 0.0,
      "op_accuracy": 0.1875,
      "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_qwen_structural_expansion",
      "global_step": 3,
      "stage_idx": 3
    },
    {
      "stage": "stage3_max24",
      "split": "paraphrase_L8",
      "n": 2,
      "max_steps": 24,
      "executor_accuracy": 0.0,
      "executor_target_mass": 0.0102605358697474,
      "direct_accuracy": NaN,
      "init_accuracy": 0.0,
      "op_accuracy": 0.5625,
      "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_qwen_structural_expansion",
      "global_step": 3,
      "stage_idx": 3
    },
    {
      "stage": "stage3_max24",
      "split": "paired_L8",
      "n": 2,
      "max_steps": 24,
      "executor_accuracy": 0.5,
      "executor_target_mass": 0.010261952877044678,
      "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_qwen_structural_expansion",
      "global_step": 3,
      "stage_idx": 3
    },
    {
      "stage": "stage3_max24",
      "split": "standard_L16",
      "n": 2,
      "max_steps": 24,
      "executor_accuracy": 0.0,
      "executor_target_mass": 0.010262958239763975,
      "direct_accuracy": NaN,
      "init_accuracy": 0.0,
      "op_accuracy": 0.4375,
      "arg_accuracy": 0.0625,
      "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_qwen_structural_expansion",
      "global_step": 3,
      "stage_idx": 3
    },
    {
      "stage": "stage3_max24",
      "split": "paraphrase_L16",
      "n": 2,
      "max_steps": 24,
      "executor_accuracy": 0.0,
      "executor_target_mass": 0.01025112671777606,
      "direct_accuracy": NaN,
      "init_accuracy": 0.0,
      "op_accuracy": 0.25,
      "arg_accuracy": 0.0625,
      "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_qwen_structural_expansion",
      "global_step": 3,
      "stage_idx": 3
    },
    {
      "stage": "stage3_max24",
      "split": "paired_L16",
      "n": 2,
      "max_steps": 24,
      "executor_accuracy": 0.0,
      "executor_target_mass": 0.010248606093227863,
      "direct_accuracy": NaN,
      "init_accuracy": 0.0,
      "op_accuracy": 0.0625,
      "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": 0.0,
      "executor_pair_both_correct": 0.0,
      "compiler_pair_program_consistency": 0.0,
      "compiler_pair_state_consistency": 0.0,
      "run": "smoke_qwen_structural_expansion",
      "global_step": 3,
      "stage_idx": 3
    },
    {
      "stage": "stage3_max24",
      "split": "standard_L24",
      "n": 2,
      "max_steps": 24,
      "executor_accuracy": 0.0,
      "executor_target_mass": 0.010253596119582653,
      "direct_accuracy": NaN,
      "init_accuracy": 0.0,
      "op_accuracy": 0.3541666666666667,
      "arg_accuracy": 0.0,
      "program_exact": 0.0,
      "state_accuracy": 0.020833333333333332,
      "state_all_exact": 0.0,
      "state_prefix_fraction": 0.0,
      "state_mean_correct_prefix": 0.0,
      "run": "smoke_qwen_structural_expansion",
      "global_step": 3,
      "stage_idx": 3
    },
    {
      "stage": "stage3_max24",
      "split": "paraphrase_L24",
      "n": 2,
      "max_steps": 24,
      "executor_accuracy": 0.0,
      "executor_target_mass": 0.010253993794322014,
      "direct_accuracy": NaN,
      "init_accuracy": 0.0,
      "op_accuracy": 0.2708333333333333,
      "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_qwen_structural_expansion",
      "global_step": 3,
      "stage_idx": 3
    },
    {
      "stage": "stage3_max24",
      "split": "paired_L24",
      "n": 2,
      "max_steps": 24,
      "executor_accuracy": 0.0,
      "executor_target_mass": 0.010250895284116268,
      "direct_accuracy": NaN,
      "init_accuracy": 0.0,
      "op_accuracy": 0.375,
      "arg_accuracy": 0.0,
      "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_qwen_structural_expansion",
      "global_step": 3,
      "stage_idx": 3
    }
  ],
  "checkpoints": [],
  "elapsed_sec": 6.944
}