{
  "args": {
    "model_id": "hf-internal-testing/tiny-random-LlamaForCausalLM",
    "text_model_fallback": "Qwen/Qwen3-4B",
    "model_class": "causal",
    "load_in_4bit": false,
    "torch_dtype": "bf16",
    "device_map": "auto",
    "modulus": 17,
    "max_steps": 5,
    "train_min_len": 1,
    "train_max_len": 2,
    "eval_lengths": "2,4",
    "train_size": 24,
    "eval_size": 16,
    "feature_batch_size": 8,
    "train_batch_size": 8,
    "eval_batch_size": 16,
    "train_steps": 6,
    "eval_every": 3,
    "lr": 0.001,
    "weight_decay": 0.01,
    "grad_clip": 1.0,
    "head_width": 64,
    "trace_loss_weight": 1.0,
    "executor_loss_weight": 1.0,
    "variants": "direct,compiler_trace,compiler_answer_only",
    "max_length": 512,
    "seed": 7,
    "output_dir": "experiments/qwen_structured_bridge/runs/smoke_tiny",
    "checkpoint_dir": "",
    "verbose": false
  },
  "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,
    "gpu_name": "NVIDIA RTX 6000 Ada Generation",
    "gpu_vram_gb": 47.383
  },
  "hidden_dim": 16,
  "dataset": {
    "train_size": 24,
    "eval_size": 16,
    "train_lengths": [
      1,
      2
    ],
    "eval_lengths": [
      2,
      4
    ]
  },
  "variants": {
    "direct": {
      "variant": "direct",
      "train_seconds": 0.13118791580200195,
      "checkpoints": [
        "large_artifacts/qwen_structured_bridge/checkpoints/smoke_tiny/direct/checkpoint_final.pt"
      ],
      "train_log": [
        {
          "step": 1,
          "loss": 2.8703970909118652,
          "direct_loss": 2.8703970909118652,
          "len2_direct_accuracy": 0.0625,
          "len2_direct_target_mass": 0.06003369018435478,
          "len2_executor_accuracy": NaN,
          "len2_executor_target_mass": NaN,
          "len2_init_accuracy": NaN,
          "len2_op_accuracy": NaN,
          "len2_arg_accuracy": NaN,
          "len2_program_exact": NaN,
          "len4_direct_accuracy": 0.0625,
          "len4_direct_target_mass": 0.05962216109037399,
          "len4_executor_accuracy": NaN,
          "len4_executor_target_mass": NaN,
          "len4_init_accuracy": NaN,
          "len4_op_accuracy": NaN,
          "len4_arg_accuracy": NaN,
          "len4_program_exact": NaN
        },
        {
          "step": 3,
          "loss": 2.804377794265747,
          "direct_loss": 2.804377794265747,
          "len2_direct_accuracy": 0.0625,
          "len2_direct_target_mass": 0.05983434617519379,
          "len2_executor_accuracy": NaN,
          "len2_executor_target_mass": NaN,
          "len2_init_accuracy": NaN,
          "len2_op_accuracy": NaN,
          "len2_arg_accuracy": NaN,
          "len2_program_exact": NaN,
          "len4_direct_accuracy": 0.0625,
          "len4_direct_target_mass": 0.05865258350968361,
          "len4_executor_accuracy": NaN,
          "len4_executor_target_mass": NaN,
          "len4_init_accuracy": NaN,
          "len4_op_accuracy": NaN,
          "len4_arg_accuracy": NaN,
          "len4_program_exact": NaN
        },
        {
          "step": 6,
          "loss": 2.831634044647217,
          "direct_loss": 2.831634044647217,
          "len2_direct_accuracy": 0.0625,
          "len2_direct_target_mass": 0.059915900230407715,
          "len2_executor_accuracy": NaN,
          "len2_executor_target_mass": NaN,
          "len2_init_accuracy": NaN,
          "len2_op_accuracy": NaN,
          "len2_arg_accuracy": NaN,
          "len2_program_exact": NaN,
          "len4_direct_accuracy": 0.0625,
          "len4_direct_target_mass": 0.05736011266708374,
          "len4_executor_accuracy": NaN,
          "len4_executor_target_mass": NaN,
          "len4_init_accuracy": NaN,
          "len4_op_accuracy": NaN,
          "len4_arg_accuracy": NaN,
          "len4_program_exact": NaN
        }
      ],
      "final_metrics": {
        "len2": {
          "variant": "direct",
          "n": 16.0,
          "direct_accuracy": 0.0625,
          "direct_target_mass": 0.059915900230407715,
          "executor_accuracy": NaN,
          "executor_target_mass": NaN,
          "init_accuracy": NaN,
          "op_accuracy": NaN,
          "arg_accuracy": NaN,
          "program_exact": NaN
        },
        "len4": {
          "variant": "direct",
          "n": 16.0,
          "direct_accuracy": 0.0625,
          "direct_target_mass": 0.05736011266708374,
          "executor_accuracy": NaN,
          "executor_target_mass": NaN,
          "init_accuracy": NaN,
          "op_accuracy": NaN,
          "arg_accuracy": NaN,
          "program_exact": NaN
        }
      }
    },
    "compiler_trace": {
      "variant": "compiler_trace",
      "train_seconds": 0.17159557342529297,
      "checkpoints": [
        "large_artifacts/qwen_structured_bridge/checkpoints/smoke_tiny/compiler_trace/checkpoint_final.pt"
      ],
      "train_log": [
        {
          "step": 1,
          "loss": 9.594647407531738,
          "executor_loss": 2.851036787033081,
          "init_loss": 2.71956205368042,
          "op_loss": 1.1163829565048218,
          "arg_loss": 2.907665729522705,
          "len2_direct_accuracy": NaN,
          "len2_direct_target_mass": NaN,
          "len2_executor_accuracy": 0.0,
          "len2_executor_target_mass": 0.06144434213638306,
          "len2_init_accuracy": 0.125,
          "len2_op_accuracy": 0.25,
          "len2_arg_accuracy": 0.03125,
          "len2_program_exact": 0.0,
          "len4_direct_accuracy": NaN,
          "len4_direct_target_mass": NaN,
          "len4_executor_accuracy": 0.0,
          "len4_executor_target_mass": 0.05736008659005165,
          "len4_init_accuracy": 0.0625,
          "len4_op_accuracy": 0.3125,
          "len4_arg_accuracy": 0.0625,
          "len4_program_exact": 0.0
        },
        {
          "step": 3,
          "loss": 9.587385177612305,
          "executor_loss": 2.8491358757019043,
          "init_loss": 2.7522807121276855,
          "op_loss": 1.0940488576889038,
          "arg_loss": 2.8919191360473633,
          "len2_direct_accuracy": NaN,
          "len2_direct_target_mass": NaN,
          "len2_executor_accuracy": 0.0,
          "len2_executor_target_mass": 0.061524730175733566,
          "len2_init_accuracy": 0.125,
          "len2_op_accuracy": 0.21875,
          "len2_arg_accuracy": 0.03125,
          "len2_program_exact": 0.0,
          "len4_direct_accuracy": NaN,
          "len4_direct_target_mass": NaN,
          "len4_executor_accuracy": 0.0625,
          "len4_executor_target_mass": 0.05727759003639221,
          "len4_init_accuracy": 0.0625,
          "len4_op_accuracy": 0.234375,
          "len4_arg_accuracy": 0.0625,
          "len4_program_exact": 0.0
        },
        {
          "step": 6,
          "loss": 9.504006385803223,
          "executor_loss": 2.847149610519409,
          "init_loss": 2.6438663005828857,
          "op_loss": 1.0677199363708496,
          "arg_loss": 2.945270299911499,
          "len2_direct_accuracy": NaN,
          "len2_direct_target_mass": NaN,
          "len2_executor_accuracy": 0.0625,
          "len2_executor_target_mass": 0.061518967151641846,
          "len2_init_accuracy": 0.125,
          "len2_op_accuracy": 0.40625,
          "len2_arg_accuracy": 0.03125,
          "len2_program_exact": 0.0,
          "len4_direct_accuracy": NaN,
          "len4_direct_target_mass": NaN,
          "len4_executor_accuracy": 0.1875,
          "len4_executor_target_mass": 0.05723065882921219,
          "len4_init_accuracy": 0.0625,
          "len4_op_accuracy": 0.296875,
          "len4_arg_accuracy": 0.0625,
          "len4_program_exact": 0.0
        }
      ],
      "final_metrics": {
        "len2": {
          "variant": "compiler_trace",
          "n": 16.0,
          "direct_accuracy": NaN,
          "direct_target_mass": NaN,
          "executor_accuracy": 0.0625,
          "executor_target_mass": 0.061518967151641846,
          "init_accuracy": 0.125,
          "op_accuracy": 0.40625,
          "arg_accuracy": 0.03125,
          "program_exact": 0.0
        },
        "len4": {
          "variant": "compiler_trace",
          "n": 16.0,
          "direct_accuracy": NaN,
          "direct_target_mass": NaN,
          "executor_accuracy": 0.1875,
          "executor_target_mass": 0.05723065882921219,
          "init_accuracy": 0.0625,
          "op_accuracy": 0.296875,
          "arg_accuracy": 0.0625,
          "program_exact": 0.0
        }
      }
    },
    "compiler_answer_only": {
      "variant": "compiler_answer_only",
      "train_seconds": 0.0634922981262207,
      "checkpoints": [
        "large_artifacts/qwen_structured_bridge/checkpoints/smoke_tiny/compiler_answer_only/checkpoint_final.pt"
      ],
      "train_log": [
        {
          "step": 1,
          "loss": 2.8145549297332764,
          "executor_loss": 2.8145549297332764,
          "len2_direct_accuracy": NaN,
          "len2_direct_target_mass": NaN,
          "len2_executor_accuracy": 0.0,
          "len2_executor_target_mass": 0.061798516660928726,
          "len2_init_accuracy": 0.0625,
          "len2_op_accuracy": 0.25,
          "len2_arg_accuracy": 0.09375,
          "len2_program_exact": 0.0,
          "len4_direct_accuracy": NaN,
          "len4_direct_target_mass": NaN,
          "len4_executor_accuracy": 0.0625,
          "len4_executor_target_mass": 0.057299815118312836,
          "len4_init_accuracy": 0.0,
          "len4_op_accuracy": 0.3125,
          "len4_arg_accuracy": 0.109375,
          "len4_program_exact": 0.0
        },
        {
          "step": 3,
          "loss": 2.808624267578125,
          "executor_loss": 2.808624267578125,
          "len2_direct_accuracy": NaN,
          "len2_direct_target_mass": NaN,
          "len2_executor_accuracy": 0.0,
          "len2_executor_target_mass": 0.06245218962430954,
          "len2_init_accuracy": 0.0625,
          "len2_op_accuracy": 0.25,
          "len2_arg_accuracy": 0.0,
          "len2_program_exact": 0.0,
          "len4_direct_accuracy": NaN,
          "len4_direct_target_mass": NaN,
          "len4_executor_accuracy": 0.125,
          "len4_executor_target_mass": 0.05691714584827423,
          "len4_init_accuracy": 0.0,
          "len4_op_accuracy": 0.3125,
          "len4_arg_accuracy": 0.015625,
          "len4_program_exact": 0.0
        },
        {
          "step": 6,
          "loss": 2.8050990104675293,
          "executor_loss": 2.8050990104675293,
          "len2_direct_accuracy": NaN,
          "len2_direct_target_mass": NaN,
          "len2_executor_accuracy": 0.0,
          "len2_executor_target_mass": 0.06254139542579651,
          "len2_init_accuracy": 0.0,
          "len2_op_accuracy": 0.25,
          "len2_arg_accuracy": 0.0,
          "len2_program_exact": 0.0,
          "len4_direct_accuracy": NaN,
          "len4_direct_target_mass": NaN,
          "len4_executor_accuracy": 0.0625,
          "len4_executor_target_mass": 0.05683267116546631,
          "len4_init_accuracy": 0.0,
          "len4_op_accuracy": 0.3125,
          "len4_arg_accuracy": 0.0,
          "len4_program_exact": 0.0
        }
      ],
      "final_metrics": {
        "len2": {
          "variant": "compiler_answer_only",
          "n": 16.0,
          "direct_accuracy": NaN,
          "direct_target_mass": NaN,
          "executor_accuracy": 0.0,
          "executor_target_mass": 0.06254139542579651,
          "init_accuracy": 0.0,
          "op_accuracy": 0.25,
          "arg_accuracy": 0.0,
          "program_exact": 0.0
        },
        "len4": {
          "variant": "compiler_answer_only",
          "n": 16.0,
          "direct_accuracy": NaN,
          "direct_target_mass": NaN,
          "executor_accuracy": 0.0625,
          "executor_target_mass": 0.05683267116546631,
          "init_accuracy": 0.0,
          "op_accuracy": 0.3125,
          "arg_accuracy": 0.0,
          "program_exact": 0.0
        }
      }
    }
  }
}