{
  "args": {
    "model_id": "Qwen/Qwen3-4B",
    "load_in_4bit": true,
    "torch_dtype": "bf16",
    "device_map": "auto",
    "use_lora": false,
    "lora_r": 8,
    "lora_alpha": 16,
    "lora_dropout": 0.05,
    "lora_target_modules": "all-linear",
    "gradient_checkpointing": true,
    "optimizer": "adamw",
    "modulus": 97,
    "max_steps": 4,
    "train_min_len": 1,
    "train_max_len": 12,
    "curriculum_stages": "smoke:1:3:2",
    "answer_train_min_len": 1,
    "answer_train_max_len": 24,
    "eval_lengths": "2,4",
    "train_template_mode": "mixed",
    "paired_train": true,
    "paired_batches": true,
    "paired_template_modes": "standard,paraphrase",
    "pair_size": 2,
    "answer_train_template_mode": "mixed",
    "eval_template_modes": "standard,paraphrase",
    "paired_eval": true,
    "paired_eval_template_modes": "standard,paraphrase",
    "train_size": 8,
    "answer_train_size": 8,
    "eval_size": 8,
    "train_batch_size": 2,
    "eval_batch_size": 2,
    "bootstrap_steps": 0,
    "answer_steps": 0,
    "eval_every": 1,
    "stage_eval_every": 1,
    "lr": 0.0002,
    "answer_lr_scale": 0.1,
    "weight_decay": 0.0,
    "grad_clip": 1.0,
    "head_width": 64,
    "register_width": 64,
    "register_layers": 1,
    "register_heads": 4,
    "register_dropout": 0.05,
    "register_style": "bare",
    "trace_loss_weight": 1.0,
    "init_trace_loss_weight": 4.0,
    "op_trace_loss_weight": 1.0,
    "arg_trace_loss_weight": 4.0,
    "executor_loss_weight": 1.0,
    "state_loss_weight": 1.0,
    "pair_consistency_loss_weight": 0.1,
    "pair_state_consistency_loss_weight": 0.5,
    "variants": "structured_trace_state_consistency",
    "max_length": 512,
    "seed": 7,
    "train_seed": -1,
    "output_dir": "experiments/qwen_register_structured_runtime/runs/smoke_frozen_structured_state_consistency",
    "checkpoint_dir": "large_artifacts/qwen_register_structured_runtime/checkpoints/smoke_frozen_structured_state_consistency"
  },
  "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": false,
    "lora_r": 8,
    "gpu_name": "NVIDIA RTX 6000 Ada Generation",
    "gpu_vram_gb": 47.383
  },
  "hidden_dim": 2560,
  "register_count": 9,
  "dataset": {
    "train_size": 8,
    "train_examples": 16,
    "curriculum_stages": [
      {
        "name": "smoke",
        "min_len": 1,
        "max_len": 3,
        "steps": 2
      }
    ],
    "answer_train_size": 8,
    "eval_size": 8,
    "eval_lengths": [
      2,
      4
    ],
    "train_template_mode": "mixed",
    "paired_train": true,
    "paired_template_modes": [
      "standard",
      "paraphrase"
    ],
    "answer_train_template_mode": "mixed",
    "eval_template_modes": [
      "standard",
      "paraphrase"
    ],
    "paired_eval": true,
    "paired_eval_template_modes": [
      "standard",
      "paraphrase"
    ],
    "register_style": "bare"
  },
  "variants": {
    "structured_trace_state_consistency": {
      "variant": "structured_trace_state_consistency",
      "train_seconds": 8.428727388381958,
      "checkpoints": [
        "large_artifacts/qwen_register_structured_runtime/checkpoints/smoke_frozen_structured_state_consistency/structured_trace_state_consistency/frozen_backbone.txt",
        "large_artifacts/qwen_register_structured_runtime/checkpoints/smoke_frozen_structured_state_consistency/structured_trace_state_consistency/heads.pt"
      ],
      "train_log": [
        {
          "variant": "structured_trace_state_consistency",
          "stage": "structured_trace_state_consistency_smoke",
          "local_step": 1,
          "step": 1,
          "loss": 46.60894775390625,
          "trace_loss_active": true,
          "executor_loss_active": true,
          "state_loss_active": true,
          "lr_scale": 1.0,
          "executor_loss": 4.577223777770996,
          "state_loss": 4.577223777770996,
          "state_train_accuracy": 0.0,
          "pair_program_loss": 0.0023628361523151398,
          "pair_state_loss": 6.170384949655272e-07,
          "init_loss": 4.650869369506836,
          "op_loss": 1.2801340818405151,
          "arg_loss": 4.392662048339844,
          "standard_len2_direct_accuracy": NaN,
          "standard_len2_direct_target_mass": NaN,
          "standard_len2_executor_accuracy": 0.0,
          "standard_len2_executor_target_mass": 0.01026506070047617,
          "standard_len2_init_accuracy": 0.0,
          "standard_len2_op_accuracy": 0.25,
          "standard_len2_arg_accuracy": 0.0625,
          "standard_len2_program_exact": 0.0,
          "standard_len2_state_accuracy": 0.0,
          "standard_len2_state_all_exact": 0.0,
          "standard_len2_state_prefix_fraction": 0.0,
          "standard_len4_direct_accuracy": NaN,
          "standard_len4_direct_target_mass": NaN,
          "standard_len4_executor_accuracy": 0.0,
          "standard_len4_executor_target_mass": 0.010262322844937444,
          "standard_len4_init_accuracy": 0.0,
          "standard_len4_op_accuracy": 0.28125,
          "standard_len4_arg_accuracy": 0.0625,
          "standard_len4_program_exact": 0.0,
          "standard_len4_state_accuracy": 0.0,
          "standard_len4_state_all_exact": 0.0,
          "standard_len4_state_prefix_fraction": 0.0,
          "paraphrase_len2_direct_accuracy": NaN,
          "paraphrase_len2_direct_target_mass": NaN,
          "paraphrase_len2_executor_accuracy": 0.0,
          "paraphrase_len2_executor_target_mass": 0.010263415053486824,
          "paraphrase_len2_init_accuracy": 0.0,
          "paraphrase_len2_op_accuracy": 0.3125,
          "paraphrase_len2_arg_accuracy": 0.125,
          "paraphrase_len2_program_exact": 0.0,
          "paraphrase_len2_state_accuracy": 0.0,
          "paraphrase_len2_state_all_exact": 0.0,
          "paraphrase_len2_state_prefix_fraction": 0.0,
          "paraphrase_len4_direct_accuracy": NaN,
          "paraphrase_len4_direct_target_mass": NaN,
          "paraphrase_len4_executor_accuracy": 0.0,
          "paraphrase_len4_executor_target_mass": 0.01026095263659954,
          "paraphrase_len4_init_accuracy": 0.0,
          "paraphrase_len4_op_accuracy": 0.40625,
          "paraphrase_len4_arg_accuracy": 0.03125,
          "paraphrase_len4_program_exact": 0.0,
          "paraphrase_len4_state_accuracy": 0.03125,
          "paraphrase_len4_state_all_exact": 0.0,
          "paraphrase_len4_state_prefix_fraction": 0.0,
          "paired_len2_direct_accuracy": NaN,
          "paired_len2_direct_target_mass": NaN,
          "paired_len2_executor_accuracy": 0.0,
          "paired_len2_executor_target_mass": 0.010262220748700202,
          "paired_len2_init_accuracy": 0.0,
          "paired_len2_op_accuracy": 0.3125,
          "paired_len2_arg_accuracy": 0.0625,
          "paired_len2_program_exact": 0.0,
          "paired_len2_state_accuracy": 0.0,
          "paired_len2_state_all_exact": 0.0,
          "paired_len2_state_prefix_fraction": 0.0,
          "paired_len2_pair_true_answer_consistency": 1.0,
          "paired_len2_executor_pair_answer_consistency": 1.0,
          "paired_len2_executor_pair_both_correct": 0.0,
          "paired_len2_compiler_pair_program_consistency": 1.0,
          "paired_len2_compiler_pair_state_consistency": 1.0,
          "paired_len4_direct_accuracy": NaN,
          "paired_len4_direct_target_mass": NaN,
          "paired_len4_executor_accuracy": 0.0,
          "paired_len4_executor_target_mass": 0.010259047034196556,
          "paired_len4_init_accuracy": 0.0,
          "paired_len4_op_accuracy": 0.3125,
          "paired_len4_arg_accuracy": 0.03125,
          "paired_len4_program_exact": 0.0,
          "paired_len4_state_accuracy": 0.0,
          "paired_len4_state_all_exact": 0.0,
          "paired_len4_state_prefix_fraction": 0.0,
          "paired_len4_pair_true_answer_consistency": 1.0,
          "paired_len4_executor_pair_answer_consistency": 1.0,
          "paired_len4_executor_pair_both_correct": 0.0,
          "paired_len4_compiler_pair_program_consistency": 1.0,
          "paired_len4_compiler_pair_state_consistency": 1.0
        },
        {
          "variant": "structured_trace_state_consistency",
          "stage": "structured_trace_state_consistency_smoke",
          "local_step": 2,
          "step": 2,
          "loss": 46.166046142578125,
          "trace_loss_active": true,
          "executor_loss_active": true,
          "state_loss_active": true,
          "lr_scale": 1.0,
          "executor_loss": 4.579158306121826,
          "state_loss": 4.579436302185059,
          "state_train_accuracy": 0.0,
          "pair_program_loss": 0.0024667230900377035,
          "pair_state_loss": 6.794907676521689e-06,
          "init_loss": 4.490883827209473,
          "op_loss": 1.237235426902771,
          "arg_loss": 4.4516072273254395,
          "standard_len2_direct_accuracy": NaN,
          "standard_len2_direct_target_mass": NaN,
          "standard_len2_executor_accuracy": 0.0,
          "standard_len2_executor_target_mass": 0.010268166894093156,
          "standard_len2_init_accuracy": 0.0,
          "standard_len2_op_accuracy": 0.25,
          "standard_len2_arg_accuracy": 0.0625,
          "standard_len2_program_exact": 0.0,
          "standard_len2_state_accuracy": 0.0,
          "standard_len2_state_all_exact": 0.0,
          "standard_len2_state_prefix_fraction": 0.0,
          "standard_len4_direct_accuracy": NaN,
          "standard_len4_direct_target_mass": NaN,
          "standard_len4_executor_accuracy": 0.0,
          "standard_len4_executor_target_mass": 0.010264950571581721,
          "standard_len4_init_accuracy": 0.0,
          "standard_len4_op_accuracy": 0.28125,
          "standard_len4_arg_accuracy": 0.0625,
          "standard_len4_program_exact": 0.0,
          "standard_len4_state_accuracy": 0.0,
          "standard_len4_state_all_exact": 0.0,
          "standard_len4_state_prefix_fraction": 0.0,
          "paraphrase_len2_direct_accuracy": NaN,
          "paraphrase_len2_direct_target_mass": NaN,
          "paraphrase_len2_executor_accuracy": 0.0,
          "paraphrase_len2_executor_target_mass": 0.010266257217153907,
          "paraphrase_len2_init_accuracy": 0.0,
          "paraphrase_len2_op_accuracy": 0.3125,
          "paraphrase_len2_arg_accuracy": 0.125,
          "paraphrase_len2_program_exact": 0.0,
          "paraphrase_len2_state_accuracy": 0.0,
          "paraphrase_len2_state_all_exact": 0.0,
          "paraphrase_len2_state_prefix_fraction": 0.0,
          "paraphrase_len4_direct_accuracy": NaN,
          "paraphrase_len4_direct_target_mass": NaN,
          "paraphrase_len4_executor_accuracy": 0.0,
          "paraphrase_len4_executor_target_mass": 0.01026359829120338,
          "paraphrase_len4_init_accuracy": 0.0,
          "paraphrase_len4_op_accuracy": 0.40625,
          "paraphrase_len4_arg_accuracy": 0.03125,
          "paraphrase_len4_program_exact": 0.0,
          "paraphrase_len4_state_accuracy": 0.03125,
          "paraphrase_len4_state_all_exact": 0.0,
          "paraphrase_len4_state_prefix_fraction": 0.0,
          "paired_len2_direct_accuracy": NaN,
          "paired_len2_direct_target_mass": NaN,
          "paired_len2_executor_accuracy": 0.0,
          "paired_len2_executor_target_mass": 0.010265539982356131,
          "paired_len2_init_accuracy": 0.0,
          "paired_len2_op_accuracy": 0.3125,
          "paired_len2_arg_accuracy": 0.0625,
          "paired_len2_program_exact": 0.0,
          "paired_len2_state_accuracy": 0.0,
          "paired_len2_state_all_exact": 0.0,
          "paired_len2_state_prefix_fraction": 0.0,
          "paired_len2_pair_true_answer_consistency": 1.0,
          "paired_len2_executor_pair_answer_consistency": 1.0,
          "paired_len2_executor_pair_both_correct": 0.0,
          "paired_len2_compiler_pair_program_consistency": 1.0,
          "paired_len2_compiler_pair_state_consistency": 1.0,
          "paired_len4_direct_accuracy": NaN,
          "paired_len4_direct_target_mass": NaN,
          "paired_len4_executor_accuracy": 0.0,
          "paired_len4_executor_target_mass": 0.010261334711685777,
          "paired_len4_init_accuracy": 0.0,
          "paired_len4_op_accuracy": 0.3125,
          "paired_len4_arg_accuracy": 0.03125,
          "paired_len4_program_exact": 0.0,
          "paired_len4_state_accuracy": 0.0,
          "paired_len4_state_all_exact": 0.0,
          "paired_len4_state_prefix_fraction": 0.0,
          "paired_len4_pair_true_answer_consistency": 1.0,
          "paired_len4_executor_pair_answer_consistency": 1.0,
          "paired_len4_executor_pair_both_correct": 0.0,
          "paired_len4_compiler_pair_program_consistency": 1.0,
          "paired_len4_compiler_pair_state_consistency": 1.0
        }
      ],
      "final_metrics": {
        "standard_len2": {
          "variant": "structured_trace_state_consistency",
          "n": 8.0,
          "direct_accuracy": NaN,
          "direct_target_mass": NaN,
          "executor_accuracy": 0.0,
          "executor_target_mass": 0.010268166894093156,
          "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
        },
        "standard_len4": {
          "variant": "structured_trace_state_consistency",
          "n": 8.0,
          "direct_accuracy": NaN,
          "direct_target_mass": NaN,
          "executor_accuracy": 0.0,
          "executor_target_mass": 0.010264950571581721,
          "init_accuracy": 0.0,
          "op_accuracy": 0.28125,
          "arg_accuracy": 0.0625,
          "program_exact": 0.0,
          "state_accuracy": 0.0,
          "state_all_exact": 0.0,
          "state_prefix_fraction": 0.0
        },
        "paraphrase_len2": {
          "variant": "structured_trace_state_consistency",
          "n": 8.0,
          "direct_accuracy": NaN,
          "direct_target_mass": NaN,
          "executor_accuracy": 0.0,
          "executor_target_mass": 0.010266257217153907,
          "init_accuracy": 0.0,
          "op_accuracy": 0.3125,
          "arg_accuracy": 0.125,
          "program_exact": 0.0,
          "state_accuracy": 0.0,
          "state_all_exact": 0.0,
          "state_prefix_fraction": 0.0
        },
        "paraphrase_len4": {
          "variant": "structured_trace_state_consistency",
          "n": 8.0,
          "direct_accuracy": NaN,
          "direct_target_mass": NaN,
          "executor_accuracy": 0.0,
          "executor_target_mass": 0.01026359829120338,
          "init_accuracy": 0.0,
          "op_accuracy": 0.40625,
          "arg_accuracy": 0.03125,
          "program_exact": 0.0,
          "state_accuracy": 0.03125,
          "state_all_exact": 0.0,
          "state_prefix_fraction": 0.0
        },
        "paired_len2": {
          "variant": "structured_trace_state_consistency",
          "n": 16.0,
          "direct_accuracy": NaN,
          "direct_target_mass": NaN,
          "executor_accuracy": 0.0,
          "executor_target_mass": 0.010265539982356131,
          "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,
          "pair_true_answer_consistency": 1.0,
          "executor_pair_answer_consistency": 1.0,
          "executor_pair_both_correct": 0.0,
          "compiler_pair_program_consistency": 1.0,
          "compiler_pair_state_consistency": 1.0
        },
        "paired_len4": {
          "variant": "structured_trace_state_consistency",
          "n": 16.0,
          "direct_accuracy": NaN,
          "direct_target_mass": NaN,
          "executor_accuracy": 0.0,
          "executor_target_mass": 0.010261334711685777,
          "init_accuracy": 0.0,
          "op_accuracy": 0.3125,
          "arg_accuracy": 0.03125,
          "program_exact": 0.0,
          "state_accuracy": 0.0,
          "state_all_exact": 0.0,
          "state_prefix_fraction": 0.0,
          "pair_true_answer_consistency": 1.0,
          "executor_pair_answer_consistency": 1.0,
          "executor_pair_both_correct": 0.0,
          "compiler_pair_program_consistency": 1.0,
          "compiler_pair_state_consistency": 1.0
        }
      }
    }
  }
}