{
  "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": 4,
    "eval_every": 2,
    "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_numeric_spans",
    "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": 0.12075138092041016,
      "checkpoints": [
        "large_artifacts/qwen_structured_bridge/checkpoints/smoke_tiny_numeric_spans/direct/checkpoint_final.pt"
      ],
      "train_log": [
        {
          "step": 1,
          "loss": 2.8700666427612305,
          "direct_loss": 2.8700666427612305,
          "len2_direct_accuracy": 0.0625,
          "len2_direct_target_mass": 0.06002797931432724,
          "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": 0.05972643196582794,
          "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": 2,
          "loss": 2.7823381423950195,
          "direct_loss": 2.7823381423950195,
          "len2_direct_accuracy": 0.0625,
          "len2_direct_target_mass": 0.059885453432798386,
          "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": 0.059164680540561676,
          "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": 4,
          "loss": 2.935696601867676,
          "direct_loss": 2.935696601867676,
          "len2_direct_accuracy": 0.0625,
          "len2_direct_target_mass": 0.05989374220371246,
          "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": 0.05838383734226227,
          "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": 0.05989374220371246,
          "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": 0.05838383734226227,
          "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": 0.15877628326416016,
      "checkpoints": [
        "large_artifacts/qwen_structured_bridge/checkpoints/smoke_tiny_numeric_spans/compiler_trace/checkpoint_final.pt"
      ],
      "train_log": [
        {
          "step": 1,
          "loss": 9.517619132995605,
          "executor_loss": 2.861931800842285,
          "init_loss": 2.7741379737854004,
          "op_loss": 1.0715309381484985,
          "arg_loss": 2.810018301010132,
          "len2_direct_accuracy": NaN,
          "len2_direct_target_mass": NaN,
          "len2_executor_accuracy": 0.0,
          "len2_executor_target_mass": 0.06252679228782654,
          "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": 0.0,
          "len4_executor_target_mass": 0.056872107088565826,
          "len4_init_accuracy": 0.0625,
          "len4_op_accuracy": 0.296875,
          "len4_arg_accuracy": 0.109375,
          "len4_program_exact": 0.0
        },
        {
          "step": 2,
          "loss": 9.664847373962402,
          "executor_loss": 2.8613271713256836,
          "init_loss": 2.7588939666748047,
          "op_loss": 1.1423611640930176,
          "arg_loss": 2.9022650718688965,
          "len2_direct_accuracy": NaN,
          "len2_direct_target_mass": NaN,
          "len2_executor_accuracy": 0.0,
          "len2_executor_target_mass": 0.06253103166818619,
          "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": 0.0,
          "len4_executor_target_mass": 0.056866880506277084,
          "len4_init_accuracy": 0.0625,
          "len4_op_accuracy": 0.296875,
          "len4_arg_accuracy": 0.109375,
          "len4_program_exact": 0.0
        },
        {
          "step": 4,
          "loss": 9.588098526000977,
          "executor_loss": 2.860436201095581,
          "init_loss": 2.8396949768066406,
          "op_loss": 1.032028317451477,
          "arg_loss": 2.8559393882751465,
          "len2_direct_accuracy": NaN,
          "len2_direct_target_mass": NaN,
          "len2_executor_accuracy": 0.0625,
          "len2_executor_target_mass": 0.06263172626495361,
          "len2_init_accuracy": 0.1875,
          "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": 0.0,
          "len4_executor_target_mass": 0.056794993579387665,
          "len4_init_accuracy": 0.0625,
          "len4_op_accuracy": 0.296875,
          "len4_arg_accuracy": 0.109375,
          "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": 0.06263172626495361,
          "init_accuracy": 0.1875,
          "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": 0.0,
          "executor_target_mass": 0.056794993579387665,
          "init_accuracy": 0.0625,
          "op_accuracy": 0.296875,
          "arg_accuracy": 0.109375,
          "program_exact": 0.0
        }
      }
    },
    "compiler_answer_only": {
      "variant": "compiler_answer_only",
      "train_seconds": 0.053183794021606445,
      "checkpoints": [
        "large_artifacts/qwen_structured_bridge/checkpoints/smoke_tiny_numeric_spans/compiler_answer_only/checkpoint_final.pt"
      ],
      "train_log": [
        {
          "step": 1,
          "loss": 2.8494300842285156,
          "executor_loss": 2.8494300842285156,
          "len2_direct_accuracy": NaN,
          "len2_direct_target_mass": NaN,
          "len2_executor_accuracy": 0.0625,
          "len2_executor_target_mass": 0.061427511274814606,
          "len2_init_accuracy": 0.25,
          "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": 0.057691045105457306,
          "len4_init_accuracy": 0.0,
          "len4_op_accuracy": 0.3125,
          "len4_arg_accuracy": 0.015625,
          "len4_program_exact": 0.0
        },
        {
          "step": 2,
          "loss": 2.8483660221099854,
          "executor_loss": 2.8483660221099854,
          "len2_direct_accuracy": NaN,
          "len2_direct_target_mass": NaN,
          "len2_executor_accuracy": 0.125,
          "len2_executor_target_mass": 0.0612625777721405,
          "len2_init_accuracy": 0.25,
          "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": 0.05776629596948624,
          "len4_init_accuracy": 0.0,
          "len4_op_accuracy": 0.3125,
          "len4_arg_accuracy": 0.015625,
          "len4_program_exact": 0.0
        },
        {
          "step": 4,
          "loss": 2.7714293003082275,
          "executor_loss": 2.7714293003082275,
          "len2_direct_accuracy": NaN,
          "len2_direct_target_mass": NaN,
          "len2_executor_accuracy": 0.0625,
          "len2_executor_target_mass": 0.06126142665743828,
          "len2_init_accuracy": 0.25,
          "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": 0.0,
          "len4_executor_target_mass": 0.05775250867009163,
          "len4_init_accuracy": 0.0,
          "len4_op_accuracy": 0.3125,
          "len4_arg_accuracy": 0.015625,
          "len4_program_exact": 0.0
        }
      ],
      "final_metrics": {
        "len2": {
          "variant": "compiler_answer_only",
          "n": 16.0,
          "direct_accuracy": NaN,
          "direct_target_mass": NaN,
          "executor_accuracy": 0.0625,
          "executor_target_mass": 0.06126142665743828,
          "init_accuracy": 0.25,
          "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": 0.0,
          "executor_target_mass": 0.05775250867009163,
          "init_accuracy": 0.0,
          "op_accuracy": 0.3125,
          "arg_accuracy": 0.015625,
          "program_exact": 0.0
        }
      }
    }
  }
}