{
  "args": {
    "model_id": "Qwen/Qwen3.5-4B",
    "text_model_fallback": "Qwen/Qwen3-4B",
    "model_class": "auto",
    "load_in_4bit": true,
    "torch_dtype": "bf16",
    "device_map": "auto",
    "modulus": 97,
    "max_steps": 12,
    "train_min_len": 1,
    "train_max_len": 3,
    "answer_train_min_len": 1,
    "answer_train_max_len": 6,
    "eval_lengths": "3,6,12",
    "train_template_mode": "standard",
    "answer_train_template_mode": "standard",
    "eval_template_modes": "standard,paraphrase",
    "train_size": 512,
    "answer_train_size": 512,
    "eval_size": 96,
    "feature_batch_size": 8,
    "train_batch_size": 128,
    "eval_batch_size": 96,
    "bootstrap_steps": 800,
    "answer_steps": 400,
    "eval_every": 400,
    "stage_eval_every": 400,
    "lr": 0.001,
    "answer_lr_scale": 0.1,
    "weight_decay": 0.01,
    "grad_clip": 1.0,
    "head_width": 512,
    "trace_loss_weight": 1.0,
    "executor_loss_weight": 1.0,
    "variants": "compiler_trace",
    "max_length": 512,
    "feature_seq_len": 192,
    "seed": 7,
    "output_dir": "experiments/qwen_span_free_compiler/runs/pilot_qwen35_token_pool_trace",
    "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": "AutoModelForMultimodalLM",
    "peft_installed": true,
    "gpu_name": "NVIDIA RTX 6000 Ada Generation",
    "gpu_vram_gb": 47.383
  },
  "hidden_dim": 2560,
  "dataset": {
    "bootstrap_train_size": 512,
    "answer_train_size": 512,
    "eval_size": 96,
    "bootstrap_train_lengths": [
      1,
      3
    ],
    "answer_train_lengths": [
      1,
      6
    ],
    "eval_lengths": [
      3,
      6,
      12
    ],
    "train_template_mode": "standard",
    "answer_train_template_mode": "standard",
    "eval_template_modes": [
      "standard",
      "paraphrase"
    ],
    "feature_seq_len": 192
  },
  "variants": {
    "compiler_trace": {
      "variant": "compiler_trace",
      "train_seconds": 116.37196040153503,
      "checkpoints": [
        "large_artifacts/qwen_span_free_compiler/checkpoints/pilot_qwen35_token_pool_trace/compiler_trace/checkpoint_final.pt"
      ],
      "train_log": [
        {
          "stage": "trace_bootstrap",
          "local_step": 1,
          "step": 1,
          "loss": 14.848298072814941,
          "trace_loss_active": true,
          "lr_scale": 1.0,
          "executor_loss": 4.57303524017334,
          "init_loss": 4.598212242126465,
          "op_loss": 1.1040332317352295,
          "arg_loss": 4.57301664352417,
          "standard_len3_direct_accuracy": NaN,
          "standard_len3_direct_target_mass": NaN,
          "standard_len3_executor_accuracy": 0.020833333333333332,
          "standard_len3_executor_target_mass": 0.010284019634127617,
          "standard_len3_init_accuracy": 0.03125,
          "standard_len3_op_accuracy": 0.3298611111111111,
          "standard_len3_arg_accuracy": 0.034722222222222224,
          "standard_len3_program_exact": 0.0,
          "standard_len6_direct_accuracy": NaN,
          "standard_len6_direct_target_mass": NaN,
          "standard_len6_executor_accuracy": 0.0,
          "standard_len6_executor_target_mass": 0.01031864620745182,
          "standard_len6_init_accuracy": 0.0,
          "standard_len6_op_accuracy": 0.3072916666666667,
          "standard_len6_arg_accuracy": 0.06076388888888889,
          "standard_len6_program_exact": 0.0,
          "standard_len12_direct_accuracy": NaN,
          "standard_len12_direct_target_mass": NaN,
          "standard_len12_executor_accuracy": 0.0,
          "standard_len12_executor_target_mass": 0.010309107601642609,
          "standard_len12_init_accuracy": 0.020833333333333332,
          "standard_len12_op_accuracy": 0.3428819444444444,
          "standard_len12_arg_accuracy": 0.033854166666666664,
          "standard_len12_program_exact": 0.0,
          "paraphrase_len3_direct_accuracy": NaN,
          "paraphrase_len3_direct_target_mass": NaN,
          "paraphrase_len3_executor_accuracy": 0.0,
          "paraphrase_len3_executor_target_mass": 0.010394290089607239,
          "paraphrase_len3_init_accuracy": 0.020833333333333332,
          "paraphrase_len3_op_accuracy": 0.3333333333333333,
          "paraphrase_len3_arg_accuracy": 0.04513888888888889,
          "paraphrase_len3_program_exact": 0.0,
          "paraphrase_len6_direct_accuracy": NaN,
          "paraphrase_len6_direct_target_mass": NaN,
          "paraphrase_len6_executor_accuracy": 0.020833333333333332,
          "paraphrase_len6_executor_target_mass": 0.010297652333974838,
          "paraphrase_len6_init_accuracy": 0.0,
          "paraphrase_len6_op_accuracy": 0.3107638888888889,
          "paraphrase_len6_arg_accuracy": 0.050347222222222224,
          "paraphrase_len6_program_exact": 0.0,
          "paraphrase_len12_direct_accuracy": NaN,
          "paraphrase_len12_direct_target_mass": NaN,
          "paraphrase_len12_executor_accuracy": 0.03125,
          "paraphrase_len12_executor_target_mass": 0.010309184590975443,
          "paraphrase_len12_init_accuracy": 0.0,
          "paraphrase_len12_op_accuracy": 0.3229166666666667,
          "paraphrase_len12_arg_accuracy": 0.05295138888888889,
          "paraphrase_len12_program_exact": 0.0
        },
        {
          "stage": "trace_bootstrap",
          "local_step": 400,
          "step": 400,
          "loss": 10.628265380859375,
          "trace_loss_active": true,
          "lr_scale": 1.0,
          "executor_loss": 4.514659881591797,
          "init_loss": 4.472977638244629,
          "op_loss": 0.5512454509735107,
          "arg_loss": 1.0893821716308594,
          "standard_len3_direct_accuracy": NaN,
          "standard_len3_direct_target_mass": NaN,
          "standard_len3_executor_accuracy": 0.010416666666666666,
          "standard_len3_executor_target_mass": 0.010353315621614456,
          "standard_len3_init_accuracy": 0.010416666666666666,
          "standard_len3_op_accuracy": 0.5138888888888888,
          "standard_len3_arg_accuracy": 0.3645833333333333,
          "standard_len3_program_exact": 0.0,
          "standard_len6_direct_accuracy": NaN,
          "standard_len6_direct_target_mass": NaN,
          "standard_len6_executor_accuracy": 0.03125,
          "standard_len6_executor_target_mass": 0.010301264002919197,
          "standard_len6_init_accuracy": 0.0,
          "standard_len6_op_accuracy": 0.453125,
          "standard_len6_arg_accuracy": 0.2378472222222222,
          "standard_len6_program_exact": 0.0,
          "standard_len12_direct_accuracy": NaN,
          "standard_len12_direct_target_mass": NaN,
          "standard_len12_executor_accuracy": 0.0,
          "standard_len12_executor_target_mass": 0.010307617485523224,
          "standard_len12_init_accuracy": 0.010416666666666666,
          "standard_len12_op_accuracy": 0.359375,
          "standard_len12_arg_accuracy": 0.14756944444444445,
          "standard_len12_program_exact": 0.0,
          "paraphrase_len3_direct_accuracy": NaN,
          "paraphrase_len3_direct_target_mass": NaN,
          "paraphrase_len3_executor_accuracy": 0.03125,
          "paraphrase_len3_executor_target_mass": 0.010289985065658888,
          "paraphrase_len3_init_accuracy": 0.0,
          "paraphrase_len3_op_accuracy": 0.3888888888888889,
          "paraphrase_len3_arg_accuracy": 0.07291666666666667,
          "paraphrase_len3_program_exact": 0.0,
          "paraphrase_len6_direct_accuracy": NaN,
          "paraphrase_len6_direct_target_mass": NaN,
          "paraphrase_len6_executor_accuracy": 0.020833333333333332,
          "paraphrase_len6_executor_target_mass": 0.010309673845767975,
          "paraphrase_len6_init_accuracy": 0.010416666666666666,
          "paraphrase_len6_op_accuracy": 0.3923611111111111,
          "paraphrase_len6_arg_accuracy": 0.09027777777777778,
          "paraphrase_len6_program_exact": 0.0,
          "paraphrase_len12_direct_accuracy": NaN,
          "paraphrase_len12_direct_target_mass": NaN,
          "paraphrase_len12_executor_accuracy": 0.010416666666666666,
          "paraphrase_len12_executor_target_mass": 0.010308854902784029,
          "paraphrase_len12_init_accuracy": 0.010416666666666666,
          "paraphrase_len12_op_accuracy": 0.4001736111111111,
          "paraphrase_len12_arg_accuracy": 0.08854166666666667,
          "paraphrase_len12_program_exact": 0.0
        },
        {
          "stage": "trace_bootstrap",
          "local_step": 800,
          "step": 800,
          "loss": 10.43506908416748,
          "trace_loss_active": true,
          "lr_scale": 1.0,
          "executor_loss": 4.523439407348633,
          "init_loss": 4.519508361816406,
          "op_loss": 0.5790759325027466,
          "arg_loss": 0.8130455017089844,
          "standard_len3_direct_accuracy": NaN,
          "standard_len3_direct_target_mass": NaN,
          "standard_len3_executor_accuracy": 0.010416666666666666,
          "standard_len3_executor_target_mass": 0.010331993301709494,
          "standard_len3_init_accuracy": 0.010416666666666666,
          "standard_len3_op_accuracy": 0.5486111111111112,
          "standard_len3_arg_accuracy": 0.3680555555555556,
          "standard_len3_program_exact": 0.0,
          "standard_len6_direct_accuracy": NaN,
          "standard_len6_direct_target_mass": NaN,
          "standard_len6_executor_accuracy": 0.010416666666666666,
          "standard_len6_executor_target_mass": 0.010321679214636484,
          "standard_len6_init_accuracy": 0.0,
          "standard_len6_op_accuracy": 0.4704861111111111,
          "standard_len6_arg_accuracy": 0.24131944444444445,
          "standard_len6_program_exact": 0.0,
          "standard_len12_direct_accuracy": NaN,
          "standard_len12_direct_target_mass": NaN,
          "standard_len12_executor_accuracy": 0.0,
          "standard_len12_executor_target_mass": 0.010308845589558283,
          "standard_len12_init_accuracy": 0.010416666666666666,
          "standard_len12_op_accuracy": 0.4157986111111111,
          "standard_len12_arg_accuracy": 0.1684027777777778,
          "standard_len12_program_exact": 0.0,
          "paraphrase_len3_direct_accuracy": NaN,
          "paraphrase_len3_direct_target_mass": NaN,
          "paraphrase_len3_executor_accuracy": 0.010416666666666666,
          "paraphrase_len3_executor_target_mass": 0.010315950959920883,
          "paraphrase_len3_init_accuracy": 0.020833333333333332,
          "paraphrase_len3_op_accuracy": 0.375,
          "paraphrase_len3_arg_accuracy": 0.04861111111111111,
          "paraphrase_len3_program_exact": 0.0,
          "paraphrase_len6_direct_accuracy": NaN,
          "paraphrase_len6_direct_target_mass": NaN,
          "paraphrase_len6_executor_accuracy": 0.010416666666666666,
          "paraphrase_len6_executor_target_mass": 0.010326884686946869,
          "paraphrase_len6_init_accuracy": 0.010416666666666666,
          "paraphrase_len6_op_accuracy": 0.3819444444444444,
          "paraphrase_len6_arg_accuracy": 0.078125,
          "paraphrase_len6_program_exact": 0.0,
          "paraphrase_len12_direct_accuracy": NaN,
          "paraphrase_len12_direct_target_mass": NaN,
          "paraphrase_len12_executor_accuracy": 0.010416666666666666,
          "paraphrase_len12_executor_target_mass": 0.010308322807153067,
          "paraphrase_len12_init_accuracy": 0.010416666666666666,
          "paraphrase_len12_op_accuracy": 0.3845486111111111,
          "paraphrase_len12_arg_accuracy": 0.0763888888888889,
          "paraphrase_len12_program_exact": 0.0
        },
        {
          "stage": "trace_continuation",
          "local_step": 1,
          "step": 801,
          "loss": 15.38522720336914,
          "trace_loss_active": true,
          "lr_scale": 1.0,
          "executor_loss": 4.593114376068115,
          "init_loss": 4.65640115737915,
          "op_loss": 1.16288423538208,
          "arg_loss": 4.972827434539795,
          "standard_len3_direct_accuracy": NaN,
          "standard_len3_direct_target_mass": NaN,
          "standard_len3_executor_accuracy": 0.0,
          "standard_len3_executor_target_mass": 0.010319992899894714,
          "standard_len3_init_accuracy": 0.010416666666666666,
          "standard_len3_op_accuracy": 0.5868055555555556,
          "standard_len3_arg_accuracy": 0.3819444444444444,
          "standard_len3_program_exact": 0.0,
          "standard_len6_direct_accuracy": NaN,
          "standard_len6_direct_target_mass": NaN,
          "standard_len6_executor_accuracy": 0.010416666666666666,
          "standard_len6_executor_target_mass": 0.010290836294492086,
          "standard_len6_init_accuracy": 0.0,
          "standard_len6_op_accuracy": 0.4670138888888889,
          "standard_len6_arg_accuracy": 0.2361111111111111,
          "standard_len6_program_exact": 0.0,
          "standard_len12_direct_accuracy": NaN,
          "standard_len12_direct_target_mass": NaN,
          "standard_len12_executor_accuracy": 0.010416666666666666,
          "standard_len12_executor_target_mass": 0.010313507169485092,
          "standard_len12_init_accuracy": 0.010416666666666666,
          "standard_len12_op_accuracy": 0.4357638888888889,
          "standard_len12_arg_accuracy": 0.1657986111111111,
          "standard_len12_program_exact": 0.0,
          "paraphrase_len3_direct_accuracy": NaN,
          "paraphrase_len3_direct_target_mass": NaN,
          "paraphrase_len3_executor_accuracy": 0.010416666666666666,
          "paraphrase_len3_executor_target_mass": 0.010313938682277998,
          "paraphrase_len3_init_accuracy": 0.020833333333333332,
          "paraphrase_len3_op_accuracy": 0.3715277777777778,
          "paraphrase_len3_arg_accuracy": 0.04513888888888889,
          "paraphrase_len3_program_exact": 0.0,
          "paraphrase_len6_direct_accuracy": NaN,
          "paraphrase_len6_direct_target_mass": NaN,
          "paraphrase_len6_executor_accuracy": 0.0,
          "paraphrase_len6_executor_target_mass": 0.010329109306136766,
          "paraphrase_len6_init_accuracy": 0.010416666666666666,
          "paraphrase_len6_op_accuracy": 0.3784722222222222,
          "paraphrase_len6_arg_accuracy": 0.05555555555555555,
          "paraphrase_len6_program_exact": 0.0,
          "paraphrase_len12_direct_accuracy": NaN,
          "paraphrase_len12_direct_target_mass": NaN,
          "paraphrase_len12_executor_accuracy": 0.0,
          "paraphrase_len12_executor_target_mass": 0.010308810820182165,
          "paraphrase_len12_init_accuracy": 0.010416666666666666,
          "paraphrase_len12_op_accuracy": 0.3758680555555556,
          "paraphrase_len12_arg_accuracy": 0.07725694444444445,
          "paraphrase_len12_program_exact": 0.0
        },
        {
          "stage": "trace_continuation",
          "local_step": 400,
          "step": 1200,
          "loss": 11.231194496154785,
          "trace_loss_active": true,
          "lr_scale": 1.0,
          "executor_loss": 4.523499488830566,
          "init_loss": 4.468802452087402,
          "op_loss": 0.6626518368721008,
          "arg_loss": 1.5762405395507812,
          "standard_len3_direct_accuracy": NaN,
          "standard_len3_direct_target_mass": NaN,
          "standard_len3_executor_accuracy": 0.010416666666666666,
          "standard_len3_executor_target_mass": 0.01031503826379776,
          "standard_len3_init_accuracy": 0.010416666666666666,
          "standard_len3_op_accuracy": 0.5590277777777778,
          "standard_len3_arg_accuracy": 0.4166666666666667,
          "standard_len3_program_exact": 0.0,
          "standard_len6_direct_accuracy": NaN,
          "standard_len6_direct_target_mass": NaN,
          "standard_len6_executor_accuracy": 0.0,
          "standard_len6_executor_target_mass": 0.010310813163717588,
          "standard_len6_init_accuracy": 0.03125,
          "standard_len6_op_accuracy": 0.5642361111111112,
          "standard_len6_arg_accuracy": 0.3541666666666667,
          "standard_len6_program_exact": 0.0,
          "standard_len12_direct_accuracy": NaN,
          "standard_len12_direct_target_mass": NaN,
          "standard_len12_executor_accuracy": 0.03125,
          "standard_len12_executor_target_mass": 0.010309343536694845,
          "standard_len12_init_accuracy": 0.020833333333333332,
          "standard_len12_op_accuracy": 0.4114583333333333,
          "standard_len12_arg_accuracy": 0.1501736111111111,
          "standard_len12_program_exact": 0.0,
          "paraphrase_len3_direct_accuracy": NaN,
          "paraphrase_len3_direct_target_mass": NaN,
          "paraphrase_len3_executor_accuracy": 0.03125,
          "paraphrase_len3_executor_target_mass": 0.010250377158323923,
          "paraphrase_len3_init_accuracy": 0.010416666666666666,
          "paraphrase_len3_op_accuracy": 0.4027777777777778,
          "paraphrase_len3_arg_accuracy": 0.041666666666666664,
          "paraphrase_len3_program_exact": 0.0,
          "paraphrase_len6_direct_accuracy": NaN,
          "paraphrase_len6_direct_target_mass": NaN,
          "paraphrase_len6_executor_accuracy": 0.010416666666666666,
          "paraphrase_len6_executor_target_mass": 0.010308517764012018,
          "paraphrase_len6_init_accuracy": 0.020833333333333332,
          "paraphrase_len6_op_accuracy": 0.3871527777777778,
          "paraphrase_len6_arg_accuracy": 0.09375,
          "paraphrase_len6_program_exact": 0.0,
          "paraphrase_len12_direct_accuracy": NaN,
          "paraphrase_len12_direct_target_mass": NaN,
          "paraphrase_len12_executor_accuracy": 0.03125,
          "paraphrase_len12_executor_target_mass": 0.010309292624394098,
          "paraphrase_len12_init_accuracy": 0.010416666666666666,
          "paraphrase_len12_op_accuracy": 0.3611111111111111,
          "paraphrase_len12_arg_accuracy": 0.06163194444444445,
          "paraphrase_len12_program_exact": 0.0
        }
      ],
      "final_metrics": {
        "standard_len3": {
          "variant": "compiler_trace",
          "n": 96.0,
          "direct_accuracy": NaN,
          "direct_target_mass": NaN,
          "executor_accuracy": 0.010416666666666666,
          "executor_target_mass": 0.01031503826379776,
          "init_accuracy": 0.010416666666666666,
          "op_accuracy": 0.5590277777777778,
          "arg_accuracy": 0.4166666666666667,
          "program_exact": 0.0
        },
        "standard_len6": {
          "variant": "compiler_trace",
          "n": 96.0,
          "direct_accuracy": NaN,
          "direct_target_mass": NaN,
          "executor_accuracy": 0.0,
          "executor_target_mass": 0.010310813163717588,
          "init_accuracy": 0.03125,
          "op_accuracy": 0.5642361111111112,
          "arg_accuracy": 0.3541666666666667,
          "program_exact": 0.0
        },
        "standard_len12": {
          "variant": "compiler_trace",
          "n": 96.0,
          "direct_accuracy": NaN,
          "direct_target_mass": NaN,
          "executor_accuracy": 0.03125,
          "executor_target_mass": 0.010309343536694845,
          "init_accuracy": 0.020833333333333332,
          "op_accuracy": 0.4114583333333333,
          "arg_accuracy": 0.1501736111111111,
          "program_exact": 0.0
        },
        "paraphrase_len3": {
          "variant": "compiler_trace",
          "n": 96.0,
          "direct_accuracy": NaN,
          "direct_target_mass": NaN,
          "executor_accuracy": 0.03125,
          "executor_target_mass": 0.010250377158323923,
          "init_accuracy": 0.010416666666666666,
          "op_accuracy": 0.4027777777777778,
          "arg_accuracy": 0.041666666666666664,
          "program_exact": 0.0
        },
        "paraphrase_len6": {
          "variant": "compiler_trace",
          "n": 96.0,
          "direct_accuracy": NaN,
          "direct_target_mass": NaN,
          "executor_accuracy": 0.010416666666666666,
          "executor_target_mass": 0.010308517764012018,
          "init_accuracy": 0.020833333333333332,
          "op_accuracy": 0.3871527777777778,
          "arg_accuracy": 0.09375,
          "program_exact": 0.0
        },
        "paraphrase_len12": {
          "variant": "compiler_trace",
          "n": 96.0,
          "direct_accuracy": NaN,
          "direct_target_mass": NaN,
          "executor_accuracy": 0.03125,
          "executor_target_mass": 0.010309292624394098,
          "init_accuracy": 0.010416666666666666,
          "op_accuracy": 0.3611111111111111,
          "arg_accuracy": 0.06163194444444445,
          "program_exact": 0.0
        }
      }
    }
  }
}