{
  "args": {
    "model_id": "hf-internal-testing/tiny-random-LlamaForCausalLM",
    "text_model_fallback": "Qwen/Qwen3-4B",
    "model_class": "causal",
    "load_in_4bit": false,
    "torch_dtype": "fp32",
    "device_map": "auto",
    "modulus": 17,
    "max_steps": 6,
    "train_min_len": 1,
    "train_max_len": 2,
    "answer_train_min_len": 1,
    "answer_train_max_len": 4,
    "eval_lengths": "2,4",
    "train_template_mode": "standard",
    "answer_train_template_mode": "standard",
    "eval_template_modes": "standard,paraphrase",
    "train_size": 24,
    "answer_train_size": 24,
    "eval_size": 12,
    "feature_batch_size": 8,
    "train_batch_size": 8,
    "eval_batch_size": 16,
    "bootstrap_steps": 4,
    "answer_steps": 4,
    "eval_every": 4,
    "stage_eval_every": 2,
    "lr": 0.001,
    "answer_lr_scale": 0.1,
    "weight_decay": 0.01,
    "grad_clip": 1.0,
    "head_width": 64,
    "rank_temperature": 0.25,
    "trace_loss_weight": 1.0,
    "init_trace_loss_weight": 1.0,
    "op_trace_loss_weight": 1.0,
    "arg_trace_loss_weight": 1.0,
    "selection_loss_weight": 1.0,
    "role_pos_weight": 10.0,
    "executor_loss_weight": 1.0,
    "variants": "direct,compiler_trace,compiler_answer_only,compiler_trace_then_answer",
    "max_length": 512,
    "feature_seq_len": 128,
    "seed": 7,
    "output_dir": "experiments/qwen_shared_parser_compiler/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": {
    "bootstrap_train_size": 24,
    "answer_train_size": 24,
    "eval_size": 12,
    "bootstrap_train_lengths": [
      1,
      2
    ],
    "answer_train_lengths": [
      1,
      4
    ],
    "eval_lengths": [
      2,
      4
    ],
    "train_template_mode": "standard",
    "answer_train_template_mode": "standard",
    "eval_template_modes": [
      "standard",
      "paraphrase"
    ],
    "feature_seq_len": 128
  },
  "variants": {
    "direct": {
      "variant": "direct",
      "train_seconds": 0.21764731407165527,
      "checkpoints": [
        "large_artifacts/qwen_shared_parser_compiler/checkpoints/smoke_tiny/direct/checkpoint_final.pt"
      ],
      "train_log": [
        {
          "stage": "answer_only",
          "local_step": 1,
          "step": 1,
          "loss": 2.792112112045288,
          "trace_loss_active": false,
          "lr_scale": 1.0,
          "direct_loss": 2.792112112045288,
          "standard_len2_direct_accuracy": 0.08333333333333333,
          "standard_len2_direct_target_mass": 0.06300349036852519,
          "standard_len2_executor_accuracy": NaN,
          "standard_len2_executor_target_mass": NaN,
          "standard_len2_init_accuracy": NaN,
          "standard_len2_op_accuracy": NaN,
          "standard_len2_arg_accuracy": NaN,
          "standard_len2_program_exact": NaN,
          "standard_len4_direct_accuracy": 0.08333333333333333,
          "standard_len4_direct_target_mass": 0.05815687278906504,
          "standard_len4_executor_accuracy": NaN,
          "standard_len4_executor_target_mass": NaN,
          "standard_len4_init_accuracy": NaN,
          "standard_len4_op_accuracy": NaN,
          "standard_len4_arg_accuracy": NaN,
          "standard_len4_program_exact": NaN,
          "paraphrase_len2_direct_accuracy": 0.08333333333333333,
          "paraphrase_len2_direct_target_mass": 0.05778321623802185,
          "paraphrase_len2_executor_accuracy": NaN,
          "paraphrase_len2_executor_target_mass": NaN,
          "paraphrase_len2_init_accuracy": NaN,
          "paraphrase_len2_op_accuracy": NaN,
          "paraphrase_len2_arg_accuracy": NaN,
          "paraphrase_len2_program_exact": NaN,
          "paraphrase_len4_direct_accuracy": 0.08333333333333333,
          "paraphrase_len4_direct_target_mass": 0.05859164893627167,
          "paraphrase_len4_executor_accuracy": NaN,
          "paraphrase_len4_executor_target_mass": NaN,
          "paraphrase_len4_init_accuracy": NaN,
          "paraphrase_len4_op_accuracy": NaN,
          "paraphrase_len4_arg_accuracy": NaN,
          "paraphrase_len4_program_exact": NaN
        },
        {
          "stage": "answer_only",
          "local_step": 2,
          "step": 2,
          "loss": 2.857337474822998,
          "trace_loss_active": false,
          "lr_scale": 1.0,
          "direct_loss": 2.857337474822998,
          "standard_len2_direct_accuracy": 0.08333333333333333,
          "standard_len2_direct_target_mass": 0.06338492035865784,
          "standard_len2_executor_accuracy": NaN,
          "standard_len2_executor_target_mass": NaN,
          "standard_len2_init_accuracy": NaN,
          "standard_len2_op_accuracy": NaN,
          "standard_len2_arg_accuracy": NaN,
          "standard_len2_program_exact": NaN,
          "standard_len4_direct_accuracy": 0.08333333333333333,
          "standard_len4_direct_target_mass": 0.05797603726387024,
          "standard_len4_executor_accuracy": NaN,
          "standard_len4_executor_target_mass": NaN,
          "standard_len4_init_accuracy": NaN,
          "standard_len4_op_accuracy": NaN,
          "standard_len4_arg_accuracy": NaN,
          "standard_len4_program_exact": NaN,
          "paraphrase_len2_direct_accuracy": 0.08333333333333333,
          "paraphrase_len2_direct_target_mass": 0.05775047838687897,
          "paraphrase_len2_executor_accuracy": NaN,
          "paraphrase_len2_executor_target_mass": NaN,
          "paraphrase_len2_init_accuracy": NaN,
          "paraphrase_len2_op_accuracy": NaN,
          "paraphrase_len2_arg_accuracy": NaN,
          "paraphrase_len2_program_exact": NaN,
          "paraphrase_len4_direct_accuracy": 0.08333333333333333,
          "paraphrase_len4_direct_target_mass": 0.05893718699614207,
          "paraphrase_len4_executor_accuracy": NaN,
          "paraphrase_len4_executor_target_mass": NaN,
          "paraphrase_len4_init_accuracy": NaN,
          "paraphrase_len4_op_accuracy": NaN,
          "paraphrase_len4_arg_accuracy": NaN,
          "paraphrase_len4_program_exact": NaN
        },
        {
          "stage": "answer_only",
          "local_step": 4,
          "step": 4,
          "loss": 2.727606773376465,
          "trace_loss_active": false,
          "lr_scale": 1.0,
          "direct_loss": 2.727606773376465,
          "standard_len2_direct_accuracy": 0.08333333333333333,
          "standard_len2_direct_target_mass": 0.06387031078338623,
          "standard_len2_executor_accuracy": NaN,
          "standard_len2_executor_target_mass": NaN,
          "standard_len2_init_accuracy": NaN,
          "standard_len2_op_accuracy": NaN,
          "standard_len2_arg_accuracy": NaN,
          "standard_len2_program_exact": NaN,
          "standard_len4_direct_accuracy": 0.08333333333333333,
          "standard_len4_direct_target_mass": 0.05774283409118652,
          "standard_len4_executor_accuracy": NaN,
          "standard_len4_executor_target_mass": NaN,
          "standard_len4_init_accuracy": NaN,
          "standard_len4_op_accuracy": NaN,
          "standard_len4_arg_accuracy": NaN,
          "standard_len4_program_exact": NaN,
          "paraphrase_len2_direct_accuracy": 0.08333333333333333,
          "paraphrase_len2_direct_target_mass": 0.0575112650791804,
          "paraphrase_len2_executor_accuracy": NaN,
          "paraphrase_len2_executor_target_mass": NaN,
          "paraphrase_len2_init_accuracy": NaN,
          "paraphrase_len2_op_accuracy": NaN,
          "paraphrase_len2_arg_accuracy": NaN,
          "paraphrase_len2_program_exact": NaN,
          "paraphrase_len4_direct_accuracy": 0.08333333333333333,
          "paraphrase_len4_direct_target_mass": 0.05932555596033732,
          "paraphrase_len4_executor_accuracy": NaN,
          "paraphrase_len4_executor_target_mass": NaN,
          "paraphrase_len4_init_accuracy": NaN,
          "paraphrase_len4_op_accuracy": NaN,
          "paraphrase_len4_arg_accuracy": NaN,
          "paraphrase_len4_program_exact": NaN
        },
        {
          "stage": "answer_only",
          "local_step": 6,
          "step": 6,
          "loss": 2.7542505264282227,
          "trace_loss_active": false,
          "lr_scale": 1.0,
          "direct_loss": 2.7542505264282227,
          "standard_len2_direct_accuracy": 0.16666666666666666,
          "standard_len2_direct_target_mass": 0.06433513760566711,
          "standard_len2_executor_accuracy": NaN,
          "standard_len2_executor_target_mass": NaN,
          "standard_len2_init_accuracy": NaN,
          "standard_len2_op_accuracy": NaN,
          "standard_len2_arg_accuracy": NaN,
          "standard_len2_program_exact": NaN,
          "standard_len4_direct_accuracy": 0.0,
          "standard_len4_direct_target_mass": 0.057427167892456055,
          "standard_len4_executor_accuracy": NaN,
          "standard_len4_executor_target_mass": NaN,
          "standard_len4_init_accuracy": NaN,
          "standard_len4_op_accuracy": NaN,
          "standard_len4_arg_accuracy": NaN,
          "standard_len4_program_exact": NaN,
          "paraphrase_len2_direct_accuracy": 0.0,
          "paraphrase_len2_direct_target_mass": 0.057282855113347374,
          "paraphrase_len2_executor_accuracy": NaN,
          "paraphrase_len2_executor_target_mass": NaN,
          "paraphrase_len2_init_accuracy": NaN,
          "paraphrase_len2_op_accuracy": NaN,
          "paraphrase_len2_arg_accuracy": NaN,
          "paraphrase_len2_program_exact": NaN,
          "paraphrase_len4_direct_accuracy": 0.0,
          "paraphrase_len4_direct_target_mass": 0.059443081418673195,
          "paraphrase_len4_executor_accuracy": NaN,
          "paraphrase_len4_executor_target_mass": NaN,
          "paraphrase_len4_init_accuracy": NaN,
          "paraphrase_len4_op_accuracy": NaN,
          "paraphrase_len4_arg_accuracy": NaN,
          "paraphrase_len4_program_exact": NaN
        },
        {
          "stage": "answer_only",
          "local_step": 8,
          "step": 8,
          "loss": 2.696685791015625,
          "trace_loss_active": false,
          "lr_scale": 1.0,
          "direct_loss": 2.696685791015625,
          "standard_len2_direct_accuracy": 0.16666666666666666,
          "standard_len2_direct_target_mass": 0.06465814510981242,
          "standard_len2_executor_accuracy": NaN,
          "standard_len2_executor_target_mass": NaN,
          "standard_len2_init_accuracy": NaN,
          "standard_len2_op_accuracy": NaN,
          "standard_len2_arg_accuracy": NaN,
          "standard_len2_program_exact": NaN,
          "standard_len4_direct_accuracy": 0.0,
          "standard_len4_direct_target_mass": 0.05699672798315684,
          "standard_len4_executor_accuracy": NaN,
          "standard_len4_executor_target_mass": NaN,
          "standard_len4_init_accuracy": NaN,
          "standard_len4_op_accuracy": NaN,
          "standard_len4_arg_accuracy": NaN,
          "standard_len4_program_exact": NaN,
          "paraphrase_len2_direct_accuracy": 0.0,
          "paraphrase_len2_direct_target_mass": 0.0572198232014974,
          "paraphrase_len2_executor_accuracy": NaN,
          "paraphrase_len2_executor_target_mass": NaN,
          "paraphrase_len2_init_accuracy": NaN,
          "paraphrase_len2_op_accuracy": NaN,
          "paraphrase_len2_arg_accuracy": NaN,
          "paraphrase_len2_program_exact": NaN,
          "paraphrase_len4_direct_accuracy": 0.0,
          "paraphrase_len4_direct_target_mass": 0.059570337335268654,
          "paraphrase_len4_executor_accuracy": NaN,
          "paraphrase_len4_executor_target_mass": NaN,
          "paraphrase_len4_init_accuracy": NaN,
          "paraphrase_len4_op_accuracy": NaN,
          "paraphrase_len4_arg_accuracy": NaN,
          "paraphrase_len4_program_exact": NaN
        }
      ],
      "final_metrics": {
        "standard_len2": {
          "variant": "direct",
          "n": 12.0,
          "direct_accuracy": 0.16666666666666666,
          "direct_target_mass": 0.06465814510981242,
          "executor_accuracy": NaN,
          "executor_target_mass": NaN,
          "init_accuracy": NaN,
          "op_accuracy": NaN,
          "arg_accuracy": NaN,
          "program_exact": NaN
        },
        "standard_len4": {
          "variant": "direct",
          "n": 12.0,
          "direct_accuracy": 0.0,
          "direct_target_mass": 0.05699672798315684,
          "executor_accuracy": NaN,
          "executor_target_mass": NaN,
          "init_accuracy": NaN,
          "op_accuracy": NaN,
          "arg_accuracy": NaN,
          "program_exact": NaN
        },
        "paraphrase_len2": {
          "variant": "direct",
          "n": 12.0,
          "direct_accuracy": 0.0,
          "direct_target_mass": 0.0572198232014974,
          "executor_accuracy": NaN,
          "executor_target_mass": NaN,
          "init_accuracy": NaN,
          "op_accuracy": NaN,
          "arg_accuracy": NaN,
          "program_exact": NaN
        },
        "paraphrase_len4": {
          "variant": "direct",
          "n": 12.0,
          "direct_accuracy": 0.0,
          "direct_target_mass": 0.059570337335268654,
          "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": 0.9995894432067871,
      "checkpoints": [
        "large_artifacts/qwen_shared_parser_compiler/checkpoints/smoke_tiny/compiler_trace/checkpoint_final.pt"
      ],
      "train_log": [
        {
          "stage": "trace_bootstrap",
          "local_step": 1,
          "step": 1,
          "loss": 1468.8927001953125,
          "trace_loss_active": true,
          "lr_scale": 1.0,
          "executor_loss": 2.811821222305298,
          "init_loss": 2.8418917655944824,
          "op_loss": 1.1227813959121704,
          "arg_loss": 2.89263916015625,
          "init_attn_loss": 3.8276100158691406,
          "op_attn_loss": 732.2246704101562,
          "arg_attn_loss": 722.2877807617188,
          "role_loss": 0.88340163230896,
          "standard_len2_direct_accuracy": NaN,
          "standard_len2_direct_target_mass": NaN,
          "standard_len2_executor_accuracy": 0.16666666666666666,
          "standard_len2_executor_target_mass": 0.061441317200660706,
          "standard_len2_init_accuracy": 0.08333333333333333,
          "standard_len2_op_accuracy": 0.25,
          "standard_len2_arg_accuracy": 0.08333333333333333,
          "standard_len2_program_exact": 0.0,
          "standard_len4_direct_accuracy": NaN,
          "standard_len4_direct_target_mass": NaN,
          "standard_len4_executor_accuracy": 0.08333333333333333,
          "standard_len4_executor_target_mass": 0.057438597083091736,
          "standard_len4_init_accuracy": 0.08333333333333333,
          "standard_len4_op_accuracy": 0.2916666666666667,
          "standard_len4_arg_accuracy": 0.041666666666666664,
          "standard_len4_program_exact": 0.0,
          "paraphrase_len2_direct_accuracy": NaN,
          "paraphrase_len2_direct_target_mass": NaN,
          "paraphrase_len2_executor_accuracy": 0.0,
          "paraphrase_len2_executor_target_mass": 0.05931768814722697,
          "paraphrase_len2_init_accuracy": 0.0,
          "paraphrase_len2_op_accuracy": 0.25,
          "paraphrase_len2_arg_accuracy": 0.16666666666666666,
          "paraphrase_len2_program_exact": 0.0,
          "paraphrase_len4_direct_accuracy": NaN,
          "paraphrase_len4_direct_target_mass": NaN,
          "paraphrase_len4_executor_accuracy": 0.08333333333333333,
          "paraphrase_len4_executor_target_mass": 0.05944619576136271,
          "paraphrase_len4_init_accuracy": 0.0,
          "paraphrase_len4_op_accuracy": 0.25,
          "paraphrase_len4_arg_accuracy": 0.08333333333333333,
          "paraphrase_len4_program_exact": 0.0
        },
        {
          "stage": "trace_bootstrap",
          "local_step": 2,
          "step": 2,
          "loss": 1443.18994140625,
          "trace_loss_active": true,
          "lr_scale": 1.0,
          "executor_loss": 2.856269359588623,
          "init_loss": 2.8008594512939453,
          "op_loss": 1.0982586145401,
          "arg_loss": 2.8809123039245605,
          "init_attn_loss": 3.82033634185791,
          "op_attn_loss": 710.001953125,
          "arg_attn_loss": 718.8421020507812,
          "role_loss": 0.8893062472343445,
          "standard_len2_direct_accuracy": NaN,
          "standard_len2_direct_target_mass": NaN,
          "standard_len2_executor_accuracy": 0.16666666666666666,
          "standard_len2_executor_target_mass": 0.06145745515823364,
          "standard_len2_init_accuracy": 0.08333333333333333,
          "standard_len2_op_accuracy": 0.25,
          "standard_len2_arg_accuracy": 0.08333333333333333,
          "standard_len2_program_exact": 0.0,
          "standard_len4_direct_accuracy": NaN,
          "standard_len4_direct_target_mass": NaN,
          "standard_len4_executor_accuracy": 0.08333333333333333,
          "standard_len4_executor_target_mass": 0.05743551750977834,
          "standard_len4_init_accuracy": 0.08333333333333333,
          "standard_len4_op_accuracy": 0.2916666666666667,
          "standard_len4_arg_accuracy": 0.041666666666666664,
          "standard_len4_program_exact": 0.0,
          "paraphrase_len2_direct_accuracy": NaN,
          "paraphrase_len2_direct_target_mass": NaN,
          "paraphrase_len2_executor_accuracy": 0.0,
          "paraphrase_len2_executor_target_mass": 0.059320494532585144,
          "paraphrase_len2_init_accuracy": 0.0,
          "paraphrase_len2_op_accuracy": 0.25,
          "paraphrase_len2_arg_accuracy": 0.16666666666666666,
          "paraphrase_len2_program_exact": 0.0,
          "paraphrase_len4_direct_accuracy": NaN,
          "paraphrase_len4_direct_target_mass": NaN,
          "paraphrase_len4_executor_accuracy": 0.08333333333333333,
          "paraphrase_len4_executor_target_mass": 0.059448450803756714,
          "paraphrase_len4_init_accuracy": 0.0,
          "paraphrase_len4_op_accuracy": 0.25,
          "paraphrase_len4_arg_accuracy": 0.08333333333333333,
          "paraphrase_len4_program_exact": 0.0
        },
        {
          "stage": "trace_bootstrap",
          "local_step": 4,
          "step": 4,
          "loss": 1361.383056640625,
          "trace_loss_active": true,
          "lr_scale": 1.0,
          "executor_loss": 2.8121845722198486,
          "init_loss": 2.8039088249206543,
          "op_loss": 1.0792046785354614,
          "arg_loss": 2.8722217082977295,
          "init_attn_loss": 3.806222677230835,
          "op_attn_loss": 671.3572998046875,
          "arg_attn_loss": 675.7666015625,
          "role_loss": 0.8854259848594666,
          "standard_len2_direct_accuracy": NaN,
          "standard_len2_direct_target_mass": NaN,
          "standard_len2_executor_accuracy": 0.08333333333333333,
          "standard_len2_executor_target_mass": 0.06149453421433767,
          "standard_len2_init_accuracy": 0.08333333333333333,
          "standard_len2_op_accuracy": 0.25,
          "standard_len2_arg_accuracy": 0.25,
          "standard_len2_program_exact": 0.0,
          "standard_len4_direct_accuracy": NaN,
          "standard_len4_direct_target_mass": NaN,
          "standard_len4_executor_accuracy": 0.0,
          "standard_len4_executor_target_mass": 0.05742781857649485,
          "standard_len4_init_accuracy": 0.08333333333333333,
          "standard_len4_op_accuracy": 0.2916666666666667,
          "standard_len4_arg_accuracy": 0.041666666666666664,
          "standard_len4_program_exact": 0.0,
          "paraphrase_len2_direct_accuracy": NaN,
          "paraphrase_len2_direct_target_mass": NaN,
          "paraphrase_len2_executor_accuracy": 0.0,
          "paraphrase_len2_executor_target_mass": 0.059326618909835815,
          "paraphrase_len2_init_accuracy": 0.0,
          "paraphrase_len2_op_accuracy": 0.25,
          "paraphrase_len2_arg_accuracy": 0.16666666666666666,
          "paraphrase_len2_program_exact": 0.0,
          "paraphrase_len4_direct_accuracy": NaN,
          "paraphrase_len4_direct_target_mass": NaN,
          "paraphrase_len4_executor_accuracy": 0.08333333333333333,
          "paraphrase_len4_executor_target_mass": 0.05944867432117462,
          "paraphrase_len4_init_accuracy": 0.0,
          "paraphrase_len4_op_accuracy": 0.25,
          "paraphrase_len4_arg_accuracy": 0.08333333333333333,
          "paraphrase_len4_program_exact": 0.0
        },
        {
          "stage": "trace_continuation",
          "local_step": 1,
          "step": 5,
          "loss": 1871.473876953125,
          "trace_loss_active": true,
          "lr_scale": 1.0,
          "executor_loss": 2.8554024696350098,
          "init_loss": 2.909789562225342,
          "op_loss": 1.1200621128082275,
          "arg_loss": 2.851095199584961,
          "init_attn_loss": 3.8967175483703613,
          "op_attn_loss": 933.20703125,
          "arg_attn_loss": 923.72802734375,
          "role_loss": 0.9057943820953369,
          "standard_len2_direct_accuracy": NaN,
          "standard_len2_direct_target_mass": NaN,
          "standard_len2_executor_accuracy": 0.08333333333333333,
          "standard_len2_executor_target_mass": 0.06150594353675842,
          "standard_len2_init_accuracy": 0.08333333333333333,
          "standard_len2_op_accuracy": 0.25,
          "standard_len2_arg_accuracy": 0.25,
          "standard_len2_program_exact": 0.0,
          "standard_len4_direct_accuracy": NaN,
          "standard_len4_direct_target_mass": NaN,
          "standard_len4_executor_accuracy": 0.0,
          "standard_len4_executor_target_mass": 0.05742437640825907,
          "standard_len4_init_accuracy": 0.08333333333333333,
          "standard_len4_op_accuracy": 0.2916666666666667,
          "standard_len4_arg_accuracy": 0.041666666666666664,
          "standard_len4_program_exact": 0.0,
          "paraphrase_len2_direct_accuracy": NaN,
          "paraphrase_len2_direct_target_mass": NaN,
          "paraphrase_len2_executor_accuracy": 0.0,
          "paraphrase_len2_executor_target_mass": 0.059329728285471596,
          "paraphrase_len2_init_accuracy": 0.0,
          "paraphrase_len2_op_accuracy": 0.25,
          "paraphrase_len2_arg_accuracy": 0.16666666666666666,
          "paraphrase_len2_program_exact": 0.0,
          "paraphrase_len4_direct_accuracy": NaN,
          "paraphrase_len4_direct_target_mass": NaN,
          "paraphrase_len4_executor_accuracy": 0.08333333333333333,
          "paraphrase_len4_executor_target_mass": 0.05944821238517761,
          "paraphrase_len4_init_accuracy": 0.0,
          "paraphrase_len4_op_accuracy": 0.25,
          "paraphrase_len4_arg_accuracy": 0.08333333333333333,
          "paraphrase_len4_program_exact": 0.0
        },
        {
          "stage": "trace_continuation",
          "local_step": 2,
          "step": 6,
          "loss": 1613.206298828125,
          "trace_loss_active": true,
          "lr_scale": 1.0,
          "executor_loss": 2.810150623321533,
          "init_loss": 2.908470630645752,
          "op_loss": 1.1008774042129517,
          "arg_loss": 2.828155517578125,
          "init_attn_loss": 3.8880324363708496,
          "op_attn_loss": 789.287109375,
          "arg_attn_loss": 809.48486328125,
          "role_loss": 0.8985618948936462,
          "standard_len2_direct_accuracy": NaN,
          "standard_len2_direct_target_mass": NaN,
          "standard_len2_executor_accuracy": 0.08333333333333333,
          "standard_len2_executor_target_mass": 0.0615131805340449,
          "standard_len2_init_accuracy": 0.08333333333333333,
          "standard_len2_op_accuracy": 0.25,
          "standard_len2_arg_accuracy": 0.25,
          "standard_len2_program_exact": 0.0,
          "standard_len4_direct_accuracy": NaN,
          "standard_len4_direct_target_mass": NaN,
          "standard_len4_executor_accuracy": 0.0,
          "standard_len4_executor_target_mass": 0.05742183327674866,
          "standard_len4_init_accuracy": 0.08333333333333333,
          "standard_len4_op_accuracy": 0.2916666666666667,
          "standard_len4_arg_accuracy": 0.041666666666666664,
          "standard_len4_program_exact": 0.0,
          "paraphrase_len2_direct_accuracy": NaN,
          "paraphrase_len2_direct_target_mass": NaN,
          "paraphrase_len2_executor_accuracy": 0.0,
          "paraphrase_len2_executor_target_mass": 0.05933301647504171,
          "paraphrase_len2_init_accuracy": 0.0,
          "paraphrase_len2_op_accuracy": 0.25,
          "paraphrase_len2_arg_accuracy": 0.16666666666666666,
          "paraphrase_len2_program_exact": 0.0,
          "paraphrase_len4_direct_accuracy": NaN,
          "paraphrase_len4_direct_target_mass": NaN,
          "paraphrase_len4_executor_accuracy": 0.08333333333333333,
          "paraphrase_len4_executor_target_mass": 0.059447397788365684,
          "paraphrase_len4_init_accuracy": 0.0,
          "paraphrase_len4_op_accuracy": 0.2708333333333333,
          "paraphrase_len4_arg_accuracy": 0.08333333333333333,
          "paraphrase_len4_program_exact": 0.0
        },
        {
          "stage": "trace_continuation",
          "local_step": 4,
          "step": 8,
          "loss": 1718.9967041015625,
          "trace_loss_active": true,
          "lr_scale": 1.0,
          "executor_loss": 2.857100486755371,
          "init_loss": 2.889997720718384,
          "op_loss": 1.0969581604003906,
          "arg_loss": 2.8142592906951904,
          "init_attn_loss": 3.9663729667663574,
          "op_attn_loss": 837.5355834960938,
          "arg_attn_loss": 866.93017578125,
          "role_loss": 0.9062172174453735,
          "standard_len2_direct_accuracy": NaN,
          "standard_len2_direct_target_mass": NaN,
          "standard_len2_executor_accuracy": 0.0,
          "standard_len2_executor_target_mass": 0.06152583658695221,
          "standard_len2_init_accuracy": 0.0,
          "standard_len2_op_accuracy": 0.25,
          "standard_len2_arg_accuracy": 0.25,
          "standard_len2_program_exact": 0.0,
          "standard_len4_direct_accuracy": NaN,
          "standard_len4_direct_target_mass": NaN,
          "standard_len4_executor_accuracy": 0.0,
          "standard_len4_executor_target_mass": 0.057417675852775574,
          "standard_len4_init_accuracy": 0.0,
          "standard_len4_op_accuracy": 0.3125,
          "standard_len4_arg_accuracy": 0.041666666666666664,
          "standard_len4_program_exact": 0.0,
          "paraphrase_len2_direct_accuracy": NaN,
          "paraphrase_len2_direct_target_mass": NaN,
          "paraphrase_len2_executor_accuracy": 0.0,
          "paraphrase_len2_executor_target_mass": 0.05933907628059387,
          "paraphrase_len2_init_accuracy": 0.0,
          "paraphrase_len2_op_accuracy": 0.25,
          "paraphrase_len2_arg_accuracy": 0.16666666666666666,
          "paraphrase_len2_program_exact": 0.0,
          "paraphrase_len4_direct_accuracy": NaN,
          "paraphrase_len4_direct_target_mass": NaN,
          "paraphrase_len4_executor_accuracy": 0.08333333333333333,
          "paraphrase_len4_executor_target_mass": 0.0594467024008433,
          "paraphrase_len4_init_accuracy": 0.08333333333333333,
          "paraphrase_len4_op_accuracy": 0.2708333333333333,
          "paraphrase_len4_arg_accuracy": 0.08333333333333333,
          "paraphrase_len4_program_exact": 0.0
        }
      ],
      "final_metrics": {
        "standard_len2": {
          "variant": "compiler_trace",
          "n": 12.0,
          "direct_accuracy": NaN,
          "direct_target_mass": NaN,
          "executor_accuracy": 0.0,
          "executor_target_mass": 0.06152583658695221,
          "init_accuracy": 0.0,
          "op_accuracy": 0.25,
          "arg_accuracy": 0.25,
          "program_exact": 0.0
        },
        "standard_len4": {
          "variant": "compiler_trace",
          "n": 12.0,
          "direct_accuracy": NaN,
          "direct_target_mass": NaN,
          "executor_accuracy": 0.0,
          "executor_target_mass": 0.057417675852775574,
          "init_accuracy": 0.0,
          "op_accuracy": 0.3125,
          "arg_accuracy": 0.041666666666666664,
          "program_exact": 0.0
        },
        "paraphrase_len2": {
          "variant": "compiler_trace",
          "n": 12.0,
          "direct_accuracy": NaN,
          "direct_target_mass": NaN,
          "executor_accuracy": 0.0,
          "executor_target_mass": 0.05933907628059387,
          "init_accuracy": 0.0,
          "op_accuracy": 0.25,
          "arg_accuracy": 0.16666666666666666,
          "program_exact": 0.0
        },
        "paraphrase_len4": {
          "variant": "compiler_trace",
          "n": 12.0,
          "direct_accuracy": NaN,
          "direct_target_mass": NaN,
          "executor_accuracy": 0.08333333333333333,
          "executor_target_mass": 0.0594467024008433,
          "init_accuracy": 0.08333333333333333,
          "op_accuracy": 0.2708333333333333,
          "arg_accuracy": 0.08333333333333333,
          "program_exact": 0.0
        }
      }
    },
    "compiler_answer_only": {
      "variant": "compiler_answer_only",
      "train_seconds": 0.7780413627624512,
      "checkpoints": [
        "large_artifacts/qwen_shared_parser_compiler/checkpoints/smoke_tiny/compiler_answer_only/checkpoint_final.pt"
      ],
      "train_log": [
        {
          "stage": "answer_only",
          "local_step": 1,
          "step": 1,
          "loss": 2.811692953109741,
          "trace_loss_active": false,
          "lr_scale": 1.0,
          "executor_loss": 2.811692953109741,
          "standard_len2_direct_accuracy": NaN,
          "standard_len2_direct_target_mass": NaN,
          "standard_len2_executor_accuracy": 0.16666666666666666,
          "standard_len2_executor_target_mass": 0.061767369508743286,
          "standard_len2_init_accuracy": 0.16666666666666666,
          "standard_len2_op_accuracy": 0.16666666666666666,
          "standard_len2_arg_accuracy": 0.08333333333333333,
          "standard_len2_program_exact": 0.0,
          "standard_len4_direct_accuracy": NaN,
          "standard_len4_direct_target_mass": NaN,
          "standard_len4_executor_accuracy": 0.16666666666666666,
          "standard_len4_executor_target_mass": 0.05702662467956543,
          "standard_len4_init_accuracy": 0.0,
          "standard_len4_op_accuracy": 0.4375,
          "standard_len4_arg_accuracy": 0.020833333333333332,
          "standard_len4_program_exact": 0.0,
          "paraphrase_len2_direct_accuracy": NaN,
          "paraphrase_len2_direct_target_mass": NaN,
          "paraphrase_len2_executor_accuracy": 0.16666666666666666,
          "paraphrase_len2_executor_target_mass": 0.05952318509419759,
          "paraphrase_len2_init_accuracy": 0.08333333333333333,
          "paraphrase_len2_op_accuracy": 0.4166666666666667,
          "paraphrase_len2_arg_accuracy": 0.0,
          "paraphrase_len2_program_exact": 0.0,
          "paraphrase_len4_direct_accuracy": NaN,
          "paraphrase_len4_direct_target_mass": NaN,
          "paraphrase_len4_executor_accuracy": 0.0,
          "paraphrase_len4_executor_target_mass": 0.059540003538131714,
          "paraphrase_len4_init_accuracy": 0.08333333333333333,
          "paraphrase_len4_op_accuracy": 0.3958333333333333,
          "paraphrase_len4_arg_accuracy": 0.08333333333333333,
          "paraphrase_len4_program_exact": 0.0
        },
        {
          "stage": "answer_only",
          "local_step": 2,
          "step": 2,
          "loss": 2.859255790710449,
          "trace_loss_active": false,
          "lr_scale": 1.0,
          "executor_loss": 2.859255790710449,
          "standard_len2_direct_accuracy": NaN,
          "standard_len2_direct_target_mass": NaN,
          "standard_len2_executor_accuracy": 0.16666666666666666,
          "standard_len2_executor_target_mass": 0.061687727769215904,
          "standard_len2_init_accuracy": 0.16666666666666666,
          "standard_len2_op_accuracy": 0.16666666666666666,
          "standard_len2_arg_accuracy": 0.08333333333333333,
          "standard_len2_program_exact": 0.0,
          "standard_len4_direct_accuracy": NaN,
          "standard_len4_direct_target_mass": NaN,
          "standard_len4_executor_accuracy": 0.16666666666666666,
          "standard_len4_executor_target_mass": 0.057036628325780235,
          "standard_len4_init_accuracy": 0.0,
          "standard_len4_op_accuracy": 0.4375,
          "standard_len4_arg_accuracy": 0.020833333333333332,
          "standard_len4_program_exact": 0.0,
          "paraphrase_len2_direct_accuracy": NaN,
          "paraphrase_len2_direct_target_mass": NaN,
          "paraphrase_len2_executor_accuracy": 0.16666666666666666,
          "paraphrase_len2_executor_target_mass": 0.05952480435371399,
          "paraphrase_len2_init_accuracy": 0.08333333333333333,
          "paraphrase_len2_op_accuracy": 0.4166666666666667,
          "paraphrase_len2_arg_accuracy": 0.0,
          "paraphrase_len2_program_exact": 0.0,
          "paraphrase_len4_direct_accuracy": NaN,
          "paraphrase_len4_direct_target_mass": NaN,
          "paraphrase_len4_executor_accuracy": 0.0,
          "paraphrase_len4_executor_target_mass": 0.05953660607337952,
          "paraphrase_len4_init_accuracy": 0.08333333333333333,
          "paraphrase_len4_op_accuracy": 0.3958333333333333,
          "paraphrase_len4_arg_accuracy": 0.08333333333333333,
          "paraphrase_len4_program_exact": 0.0
        },
        {
          "stage": "answer_only",
          "local_step": 4,
          "step": 4,
          "loss": 2.857482433319092,
          "trace_loss_active": false,
          "lr_scale": 1.0,
          "executor_loss": 2.857482433319092,
          "standard_len2_direct_accuracy": NaN,
          "standard_len2_direct_target_mass": NaN,
          "standard_len2_executor_accuracy": 0.16666666666666666,
          "standard_len2_executor_target_mass": 0.06156867742538452,
          "standard_len2_init_accuracy": 0.16666666666666666,
          "standard_len2_op_accuracy": 0.16666666666666666,
          "standard_len2_arg_accuracy": 0.08333333333333333,
          "standard_len2_program_exact": 0.0,
          "standard_len4_direct_accuracy": NaN,
          "standard_len4_direct_target_mass": NaN,
          "standard_len4_executor_accuracy": 0.16666666666666666,
          "standard_len4_executor_target_mass": 0.057022412618001304,
          "standard_len4_init_accuracy": 0.0,
          "standard_len4_op_accuracy": 0.4375,
          "standard_len4_arg_accuracy": 0.020833333333333332,
          "standard_len4_program_exact": 0.0,
          "paraphrase_len2_direct_accuracy": NaN,
          "paraphrase_len2_direct_target_mass": NaN,
          "paraphrase_len2_executor_accuracy": 0.08333333333333333,
          "paraphrase_len2_executor_target_mass": 0.05953983465830485,
          "paraphrase_len2_init_accuracy": 0.08333333333333333,
          "paraphrase_len2_op_accuracy": 0.4166666666666667,
          "paraphrase_len2_arg_accuracy": 0.0,
          "paraphrase_len2_program_exact": 0.0,
          "paraphrase_len4_direct_accuracy": NaN,
          "paraphrase_len4_direct_target_mass": NaN,
          "paraphrase_len4_executor_accuracy": 0.08333333333333333,
          "paraphrase_len4_executor_target_mass": 0.059532235066095986,
          "paraphrase_len4_init_accuracy": 0.08333333333333333,
          "paraphrase_len4_op_accuracy": 0.2916666666666667,
          "paraphrase_len4_arg_accuracy": 0.08333333333333333,
          "paraphrase_len4_program_exact": 0.0
        },
        {
          "stage": "answer_only",
          "local_step": 6,
          "step": 6,
          "loss": 2.7548470497131348,
          "trace_loss_active": false,
          "lr_scale": 1.0,
          "executor_loss": 2.7548470497131348,
          "standard_len2_direct_accuracy": NaN,
          "standard_len2_direct_target_mass": NaN,
          "standard_len2_executor_accuracy": 0.16666666666666666,
          "standard_len2_executor_target_mass": 0.061461408933003746,
          "standard_len2_init_accuracy": 0.16666666666666666,
          "standard_len2_op_accuracy": 0.16666666666666666,
          "standard_len2_arg_accuracy": 0.08333333333333333,
          "standard_len2_program_exact": 0.0,
          "standard_len4_direct_accuracy": NaN,
          "standard_len4_direct_target_mass": NaN,
          "standard_len4_executor_accuracy": 0.08333333333333333,
          "standard_len4_executor_target_mass": 0.057037606835365295,
          "standard_len4_init_accuracy": 0.0,
          "standard_len4_op_accuracy": 0.4375,
          "standard_len4_arg_accuracy": 0.020833333333333332,
          "standard_len4_program_exact": 0.0,
          "paraphrase_len2_direct_accuracy": NaN,
          "paraphrase_len2_direct_target_mass": NaN,
          "paraphrase_len2_executor_accuracy": 0.16666666666666666,
          "paraphrase_len2_executor_target_mass": 0.059546430905659996,
          "paraphrase_len2_init_accuracy": 0.08333333333333333,
          "paraphrase_len2_op_accuracy": 0.4166666666666667,
          "paraphrase_len2_arg_accuracy": 0.0,
          "paraphrase_len2_program_exact": 0.0,
          "paraphrase_len4_direct_accuracy": NaN,
          "paraphrase_len4_direct_target_mass": NaN,
          "paraphrase_len4_executor_accuracy": 0.08333333333333333,
          "paraphrase_len4_executor_target_mass": 0.059528738260269165,
          "paraphrase_len4_init_accuracy": 0.08333333333333333,
          "paraphrase_len4_op_accuracy": 0.3125,
          "paraphrase_len4_arg_accuracy": 0.08333333333333333,
          "paraphrase_len4_program_exact": 0.0
        },
        {
          "stage": "answer_only",
          "local_step": 8,
          "step": 8,
          "loss": 2.8624470233917236,
          "trace_loss_active": false,
          "lr_scale": 1.0,
          "executor_loss": 2.8624470233917236,
          "standard_len2_direct_accuracy": NaN,
          "standard_len2_direct_target_mass": NaN,
          "standard_len2_executor_accuracy": 0.16666666666666666,
          "standard_len2_executor_target_mass": 0.061403170228004456,
          "standard_len2_init_accuracy": 0.16666666666666666,
          "standard_len2_op_accuracy": 0.16666666666666666,
          "standard_len2_arg_accuracy": 0.08333333333333333,
          "standard_len2_program_exact": 0.0,
          "standard_len4_direct_accuracy": NaN,
          "standard_len4_direct_target_mass": NaN,
          "standard_len4_executor_accuracy": 0.08333333333333333,
          "standard_len4_executor_target_mass": 0.05702728529771169,
          "standard_len4_init_accuracy": 0.0,
          "standard_len4_op_accuracy": 0.4375,
          "standard_len4_arg_accuracy": 0.020833333333333332,
          "standard_len4_program_exact": 0.0,
          "paraphrase_len2_direct_accuracy": NaN,
          "paraphrase_len2_direct_target_mass": NaN,
          "paraphrase_len2_executor_accuracy": 0.08333333333333333,
          "paraphrase_len2_executor_target_mass": 0.05956444640954336,
          "paraphrase_len2_init_accuracy": 0.08333333333333333,
          "paraphrase_len2_op_accuracy": 0.3333333333333333,
          "paraphrase_len2_arg_accuracy": 0.0,
          "paraphrase_len2_program_exact": 0.0,
          "paraphrase_len4_direct_accuracy": NaN,
          "paraphrase_len4_direct_target_mass": NaN,
          "paraphrase_len4_executor_accuracy": 0.08333333333333333,
          "paraphrase_len4_executor_target_mass": 0.05953227480252584,
          "paraphrase_len4_init_accuracy": 0.08333333333333333,
          "paraphrase_len4_op_accuracy": 0.3125,
          "paraphrase_len4_arg_accuracy": 0.08333333333333333,
          "paraphrase_len4_program_exact": 0.0
        }
      ],
      "final_metrics": {
        "standard_len2": {
          "variant": "compiler_answer_only",
          "n": 12.0,
          "direct_accuracy": NaN,
          "direct_target_mass": NaN,
          "executor_accuracy": 0.16666666666666666,
          "executor_target_mass": 0.061403170228004456,
          "init_accuracy": 0.16666666666666666,
          "op_accuracy": 0.16666666666666666,
          "arg_accuracy": 0.08333333333333333,
          "program_exact": 0.0
        },
        "standard_len4": {
          "variant": "compiler_answer_only",
          "n": 12.0,
          "direct_accuracy": NaN,
          "direct_target_mass": NaN,
          "executor_accuracy": 0.08333333333333333,
          "executor_target_mass": 0.05702728529771169,
          "init_accuracy": 0.0,
          "op_accuracy": 0.4375,
          "arg_accuracy": 0.020833333333333332,
          "program_exact": 0.0
        },
        "paraphrase_len2": {
          "variant": "compiler_answer_only",
          "n": 12.0,
          "direct_accuracy": NaN,
          "direct_target_mass": NaN,
          "executor_accuracy": 0.08333333333333333,
          "executor_target_mass": 0.05956444640954336,
          "init_accuracy": 0.08333333333333333,
          "op_accuracy": 0.3333333333333333,
          "arg_accuracy": 0.0,
          "program_exact": 0.0
        },
        "paraphrase_len4": {
          "variant": "compiler_answer_only",
          "n": 12.0,
          "direct_accuracy": NaN,
          "direct_target_mass": NaN,
          "executor_accuracy": 0.08333333333333333,
          "executor_target_mass": 0.05953227480252584,
          "init_accuracy": 0.08333333333333333,
          "op_accuracy": 0.3125,
          "arg_accuracy": 0.08333333333333333,
          "program_exact": 0.0
        }
      }
    },
    "compiler_trace_then_answer": {
      "variant": "compiler_trace_then_answer",
      "train_seconds": 0.7944438457489014,
      "checkpoints": [
        "large_artifacts/qwen_shared_parser_compiler/checkpoints/smoke_tiny/compiler_trace_then_answer/checkpoint_final.pt"
      ],
      "train_log": [
        {
          "stage": "trace_bootstrap",
          "local_step": 1,
          "step": 1,
          "loss": 1386.33642578125,
          "trace_loss_active": true,
          "lr_scale": 1.0,
          "executor_loss": 2.85426664352417,
          "init_loss": 2.825883626937866,
          "op_loss": 1.0834237337112427,
          "arg_loss": 2.8070297241210938,
          "init_attn_loss": 3.79679012298584,
          "op_attn_loss": 591.3115844726562,
          "arg_attn_loss": 780.7728881835938,
          "role_loss": 0.8845459818840027,
          "standard_len2_direct_accuracy": NaN,
          "standard_len2_direct_target_mass": NaN,
          "standard_len2_executor_accuracy": 0.16666666666666666,
          "standard_len2_executor_target_mass": 0.06138287981351217,
          "standard_len2_init_accuracy": 0.08333333333333333,
          "standard_len2_op_accuracy": 0.2916666666666667,
          "standard_len2_arg_accuracy": 0.08333333333333333,
          "standard_len2_program_exact": 0.0,
          "standard_len4_direct_accuracy": NaN,
          "standard_len4_direct_target_mass": NaN,
          "standard_len4_executor_accuracy": 0.0,
          "standard_len4_executor_target_mass": 0.05730059742927551,
          "standard_len4_init_accuracy": 0.0,
          "standard_len4_op_accuracy": 0.4375,
          "standard_len4_arg_accuracy": 0.041666666666666664,
          "standard_len4_program_exact": 0.0,
          "paraphrase_len2_direct_accuracy": NaN,
          "paraphrase_len2_direct_target_mass": NaN,
          "paraphrase_len2_executor_accuracy": 0.16666666666666666,
          "paraphrase_len2_executor_target_mass": 0.05947426954905192,
          "paraphrase_len2_init_accuracy": 0.0,
          "paraphrase_len2_op_accuracy": 0.4166666666666667,
          "paraphrase_len2_arg_accuracy": 0.16666666666666666,
          "paraphrase_len2_program_exact": 0.0,
          "paraphrase_len4_direct_accuracy": NaN,
          "paraphrase_len4_direct_target_mass": NaN,
          "paraphrase_len4_executor_accuracy": 0.08333333333333333,
          "paraphrase_len4_executor_target_mass": 0.0594925582408905,
          "paraphrase_len4_init_accuracy": 0.0,
          "paraphrase_len4_op_accuracy": 0.3333333333333333,
          "paraphrase_len4_arg_accuracy": 0.020833333333333332,
          "paraphrase_len4_program_exact": 0.0
        },
        {
          "stage": "trace_bootstrap",
          "local_step": 2,
          "step": 2,
          "loss": 1474.0084228515625,
          "trace_loss_active": true,
          "lr_scale": 1.0,
          "executor_loss": 2.813387632369995,
          "init_loss": 2.8266940116882324,
          "op_loss": 1.1433862447738647,
          "arg_loss": 2.8160202503204346,
          "init_attn_loss": 3.9021098613739014,
          "op_attn_loss": 640.9571533203125,
          "arg_attn_loss": 818.6470336914062,
          "role_loss": 0.9026117920875549,
          "standard_len2_direct_accuracy": NaN,
          "standard_len2_direct_target_mass": NaN,
          "standard_len2_executor_accuracy": 0.16666666666666666,
          "standard_len2_executor_target_mass": 0.06137949228286743,
          "standard_len2_init_accuracy": 0.08333333333333333,
          "standard_len2_op_accuracy": 0.2916666666666667,
          "standard_len2_arg_accuracy": 0.08333333333333333,
          "standard_len2_program_exact": 0.0,
          "standard_len4_direct_accuracy": NaN,
          "standard_len4_direct_target_mass": NaN,
          "standard_len4_executor_accuracy": 0.0,
          "standard_len4_executor_target_mass": 0.057307456930478416,
          "standard_len4_init_accuracy": 0.0,
          "standard_len4_op_accuracy": 0.4375,
          "standard_len4_arg_accuracy": 0.041666666666666664,
          "standard_len4_program_exact": 0.0,
          "paraphrase_len2_direct_accuracy": NaN,
          "paraphrase_len2_direct_target_mass": NaN,
          "paraphrase_len2_executor_accuracy": 0.16666666666666666,
          "paraphrase_len2_executor_target_mass": 0.05947385231653849,
          "paraphrase_len2_init_accuracy": 0.0,
          "paraphrase_len2_op_accuracy": 0.4166666666666667,
          "paraphrase_len2_arg_accuracy": 0.16666666666666666,
          "paraphrase_len2_program_exact": 0.0,
          "paraphrase_len4_direct_accuracy": NaN,
          "paraphrase_len4_direct_target_mass": NaN,
          "paraphrase_len4_executor_accuracy": 0.08333333333333333,
          "paraphrase_len4_executor_target_mass": 0.05949180324872335,
          "paraphrase_len4_init_accuracy": 0.0,
          "paraphrase_len4_op_accuracy": 0.3333333333333333,
          "paraphrase_len4_arg_accuracy": 0.020833333333333332,
          "paraphrase_len4_program_exact": 0.0
        },
        {
          "stage": "trace_bootstrap",
          "local_step": 4,
          "step": 4,
          "loss": 1247.70166015625,
          "trace_loss_active": true,
          "lr_scale": 1.0,
          "executor_loss": 2.852912425994873,
          "init_loss": 2.8112738132476807,
          "op_loss": 1.128771424293518,
          "arg_loss": 2.7943077087402344,
          "init_attn_loss": 3.730898380279541,
          "op_attn_loss": 533.276123046875,
          "arg_attn_loss": 700.2491455078125,
          "role_loss": 0.8582605123519897,
          "standard_len2_direct_accuracy": NaN,
          "standard_len2_direct_target_mass": NaN,
          "standard_len2_executor_accuracy": 0.0,
          "standard_len2_executor_target_mass": 0.06136858959992727,
          "standard_len2_init_accuracy": 0.08333333333333333,
          "standard_len2_op_accuracy": 0.2916666666666667,
          "standard_len2_arg_accuracy": 0.125,
          "standard_len2_program_exact": 0.0,
          "standard_len4_direct_accuracy": NaN,
          "standard_len4_direct_target_mass": NaN,
          "standard_len4_executor_accuracy": 0.08333333333333333,
          "standard_len4_executor_target_mass": 0.05732111632823944,
          "standard_len4_init_accuracy": 0.0,
          "standard_len4_op_accuracy": 0.4375,
          "standard_len4_arg_accuracy": 0.020833333333333332,
          "standard_len4_program_exact": 0.0,
          "paraphrase_len2_direct_accuracy": NaN,
          "paraphrase_len2_direct_target_mass": NaN,
          "paraphrase_len2_executor_accuracy": 0.25,
          "paraphrase_len2_executor_target_mass": 0.059473246335983276,
          "paraphrase_len2_init_accuracy": 0.0,
          "paraphrase_len2_op_accuracy": 0.4166666666666667,
          "paraphrase_len2_arg_accuracy": 0.16666666666666666,
          "paraphrase_len2_program_exact": 0.0,
          "paraphrase_len4_direct_accuracy": NaN,
          "paraphrase_len4_direct_target_mass": NaN,
          "paraphrase_len4_executor_accuracy": 0.0,
          "paraphrase_len4_executor_target_mass": 0.05949115256468455,
          "paraphrase_len4_init_accuracy": 0.0,
          "paraphrase_len4_op_accuracy": 0.3333333333333333,
          "paraphrase_len4_arg_accuracy": 0.0625,
          "paraphrase_len4_program_exact": 0.0
        },
        {
          "stage": "answer_retention",
          "local_step": 1,
          "step": 5,
          "loss": 2.8590030670166016,
          "trace_loss_active": false,
          "lr_scale": 1.0,
          "executor_loss": 2.8590030670166016,
          "standard_len2_direct_accuracy": NaN,
          "standard_len2_direct_target_mass": NaN,
          "standard_len2_executor_accuracy": 0.0,
          "standard_len2_executor_target_mass": 0.06133706867694855,
          "standard_len2_init_accuracy": 0.08333333333333333,
          "standard_len2_op_accuracy": 0.2916666666666667,
          "standard_len2_arg_accuracy": 0.125,
          "standard_len2_program_exact": 0.0,
          "standard_len4_direct_accuracy": NaN,
          "standard_len4_direct_target_mass": NaN,
          "standard_len4_executor_accuracy": 0.0,
          "standard_len4_executor_target_mass": 0.05733901262283325,
          "standard_len4_init_accuracy": 0.0,
          "standard_len4_op_accuracy": 0.4375,
          "standard_len4_arg_accuracy": 0.020833333333333332,
          "standard_len4_program_exact": 0.0,
          "paraphrase_len2_direct_accuracy": NaN,
          "paraphrase_len2_direct_target_mass": NaN,
          "paraphrase_len2_executor_accuracy": 0.25,
          "paraphrase_len2_executor_target_mass": 0.05946762363115946,
          "paraphrase_len2_init_accuracy": 0.0,
          "paraphrase_len2_op_accuracy": 0.4166666666666667,
          "paraphrase_len2_arg_accuracy": 0.16666666666666666,
          "paraphrase_len2_program_exact": 0.0,
          "paraphrase_len4_direct_accuracy": NaN,
          "paraphrase_len4_direct_target_mass": NaN,
          "paraphrase_len4_executor_accuracy": 0.0,
          "paraphrase_len4_executor_target_mass": 0.05948453644911448,
          "paraphrase_len4_init_accuracy": 0.0,
          "paraphrase_len4_op_accuracy": 0.3333333333333333,
          "paraphrase_len4_arg_accuracy": 0.0625,
          "paraphrase_len4_program_exact": 0.0
        },
        {
          "stage": "answer_retention",
          "local_step": 2,
          "step": 6,
          "loss": 2.853578567504883,
          "trace_loss_active": false,
          "lr_scale": 1.0,
          "executor_loss": 2.853578567504883,
          "standard_len2_direct_accuracy": NaN,
          "standard_len2_direct_target_mass": NaN,
          "standard_len2_executor_accuracy": 0.0,
          "standard_len2_executor_target_mass": 0.06129424770673116,
          "standard_len2_init_accuracy": 0.08333333333333333,
          "standard_len2_op_accuracy": 0.2916666666666667,
          "standard_len2_arg_accuracy": 0.125,
          "standard_len2_program_exact": 0.0,
          "standard_len4_direct_accuracy": NaN,
          "standard_len4_direct_target_mass": NaN,
          "standard_len4_executor_accuracy": 0.0,
          "standard_len4_executor_target_mass": 0.05735814571380615,
          "standard_len4_init_accuracy": 0.0,
          "standard_len4_op_accuracy": 0.4375,
          "standard_len4_arg_accuracy": 0.020833333333333332,
          "standard_len4_program_exact": 0.0,
          "paraphrase_len2_direct_accuracy": NaN,
          "paraphrase_len2_direct_target_mass": NaN,
          "paraphrase_len2_executor_accuracy": 0.25,
          "paraphrase_len2_executor_target_mass": 0.05946168303489685,
          "paraphrase_len2_init_accuracy": 0.0,
          "paraphrase_len2_op_accuracy": 0.4166666666666667,
          "paraphrase_len2_arg_accuracy": 0.16666666666666666,
          "paraphrase_len2_program_exact": 0.0,
          "paraphrase_len4_direct_accuracy": NaN,
          "paraphrase_len4_direct_target_mass": NaN,
          "paraphrase_len4_executor_accuracy": 0.0,
          "paraphrase_len4_executor_target_mass": 0.059476450085639954,
          "paraphrase_len4_init_accuracy": 0.0,
          "paraphrase_len4_op_accuracy": 0.3333333333333333,
          "paraphrase_len4_arg_accuracy": 0.0625,
          "paraphrase_len4_program_exact": 0.0
        },
        {
          "stage": "answer_retention",
          "local_step": 4,
          "step": 8,
          "loss": 2.8097920417785645,
          "trace_loss_active": false,
          "lr_scale": 1.0,
          "executor_loss": 2.8097920417785645,
          "standard_len2_direct_accuracy": NaN,
          "standard_len2_direct_target_mass": NaN,
          "standard_len2_executor_accuracy": 0.0,
          "standard_len2_executor_target_mass": 0.061250239610672,
          "standard_len2_init_accuracy": 0.08333333333333333,
          "standard_len2_op_accuracy": 0.2916666666666667,
          "standard_len2_arg_accuracy": 0.125,
          "standard_len2_program_exact": 0.0,
          "standard_len4_direct_accuracy": NaN,
          "standard_len4_direct_target_mass": NaN,
          "standard_len4_executor_accuracy": 0.0,
          "standard_len4_executor_target_mass": 0.05737386643886566,
          "standard_len4_init_accuracy": 0.0,
          "standard_len4_op_accuracy": 0.4375,
          "standard_len4_arg_accuracy": 0.020833333333333332,
          "standard_len4_program_exact": 0.0,
          "paraphrase_len2_direct_accuracy": NaN,
          "paraphrase_len2_direct_target_mass": NaN,
          "paraphrase_len2_executor_accuracy": 0.25,
          "paraphrase_len2_executor_target_mass": 0.05946325262387594,
          "paraphrase_len2_init_accuracy": 0.0,
          "paraphrase_len2_op_accuracy": 0.4166666666666667,
          "paraphrase_len2_arg_accuracy": 0.16666666666666666,
          "paraphrase_len2_program_exact": 0.0,
          "paraphrase_len4_direct_accuracy": NaN,
          "paraphrase_len4_direct_target_mass": NaN,
          "paraphrase_len4_executor_accuracy": 0.0,
          "paraphrase_len4_executor_target_mass": 0.05947257081667582,
          "paraphrase_len4_init_accuracy": 0.0,
          "paraphrase_len4_op_accuracy": 0.3333333333333333,
          "paraphrase_len4_arg_accuracy": 0.0625,
          "paraphrase_len4_program_exact": 0.0
        }
      ],
      "final_metrics": {
        "standard_len2": {
          "variant": "compiler_trace_then_answer",
          "n": 12.0,
          "direct_accuracy": NaN,
          "direct_target_mass": NaN,
          "executor_accuracy": 0.0,
          "executor_target_mass": 0.061250239610672,
          "init_accuracy": 0.08333333333333333,
          "op_accuracy": 0.2916666666666667,
          "arg_accuracy": 0.125,
          "program_exact": 0.0
        },
        "standard_len4": {
          "variant": "compiler_trace_then_answer",
          "n": 12.0,
          "direct_accuracy": NaN,
          "direct_target_mass": NaN,
          "executor_accuracy": 0.0,
          "executor_target_mass": 0.05737386643886566,
          "init_accuracy": 0.0,
          "op_accuracy": 0.4375,
          "arg_accuracy": 0.020833333333333332,
          "program_exact": 0.0
        },
        "paraphrase_len2": {
          "variant": "compiler_trace_then_answer",
          "n": 12.0,
          "direct_accuracy": NaN,
          "direct_target_mass": NaN,
          "executor_accuracy": 0.25,
          "executor_target_mass": 0.05946325262387594,
          "init_accuracy": 0.0,
          "op_accuracy": 0.4166666666666667,
          "arg_accuracy": 0.16666666666666666,
          "program_exact": 0.0
        },
        "paraphrase_len4": {
          "variant": "compiler_trace_then_answer",
          "n": 12.0,
          "direct_accuracy": NaN,
          "direct_target_mass": NaN,
          "executor_accuracy": 0.0,
          "executor_target_mass": 0.05947257081667582,
          "init_accuracy": 0.0,
          "op_accuracy": 0.3333333333333333,
          "arg_accuracy": 0.0625,
          "program_exact": 0.0
        }
      }
    }
  }
}