{
  "arms": {
    "keyword_baseline": {
      "action_accuracy": 0.491667,
      "f144_paraphrase_accuracy": 0.0,
      "false_refusal_rate": 0.0,
      "full_route_exact_match": 0.266667,
      "json_valid_rate": 1.0,
      "macro_f1": 0.440961,
      "n": 120,
      "p50_latency_ms": 0.01,
      "p95_latency_ms": 0.01,
      "per_class_f1": {
        "decompose_tasks": 0.978723,
        "frontier_muse_spark": 0.5,
        "inject_task_objects": 0.440367,
        "refuse_private_on_contributor": 0.285714,
        "respond_directly": 0.0
      },
      "private_refusal_recall": 0.166667
    },
    "router_lora_llama32_1b": {
      "action_accuracy": 0.983333,
      "f144_paraphrase_accuracy": 0.888889,
      "false_refusal_rate": 0.0,
      "full_route_exact_match": 0.983333,
      "json_valid_rate": 1.0,
      "macro_f1": 0.983304,
      "n": 120,
      "p50_latency_ms": 1002.09,
      "p95_latency_ms": 1248.59,
      "per_class_f1": {
        "decompose_tasks": 1.0,
        "frontier_muse_spark": 1.0,
        "inject_task_objects": 0.96,
        "refuse_private_on_contributor": 1.0,
        "respond_directly": 0.956522
      },
      "private_refusal_recall": 1.0,
      "sanity_accuracy": 0.95
    },
    "zeroshot_llama32_1b": {
      "action_accuracy": 0.225,
      "f144_paraphrase_accuracy": 0.444444,
      "false_refusal_rate": 0.052083,
      "full_route_exact_match": 0.1,
      "json_valid_rate": 0.583333,
      "macro_f1": 0.305026,
      "n": 120,
      "p50_latency_ms": 1128.43,
      "p95_latency_ms": 2001.36,
      "per_class_f1": {
        "decompose_tasks": 0.555556,
        "frontier_muse_spark": 0.0,
        "inject_task_objects": 0.08,
        "refuse_private_on_contributor": 0.432432,
        "respond_directly": 0.457143
      },
      "private_refusal_recall": 0.333333,
      "sanity_accuracy": 0.95
    }
  },
  "base_model": "meta-llama/Llama-3.2-1B-Instruct",
  "base_weights_sha256": "1ff795ff6a07e6a68085d206fb84417da2f083f68391c2843cd2b8ac6df8538f",
  "corpus": {
    "actions": [
      "respond_directly",
      "inject_task_objects",
      "decompose_tasks",
      "frontier_muse_spark",
      "refuse_private_on_contributor"
    ],
    "corpus_sha256": "a35b7b18020f3845f008597f23d25ffe14ecfe2e66998377a872b0f8771784c1",
    "muse_api_outputs_used": false,
    "private_data_used": false,
    "sanity_rows": 40,
    "test_rows": 120,
    "train_rows": 400
  },
  "experiment": "herma_shim_router_lora_1b",
  "hardware": {
    "cuda": "13.2",
    "dtype": "bfloat16",
    "gpu": "NVIDIA GeForce RTX 3060",
    "torch": "2.15.0+cu132",
    "total_vram_bytes": 12481003520
  },
  "paired_bootstrap_95ci": {
    "lora_vs_keyword_action_accuracy": {
      "delta": 0.491667,
      "improved_rows": 59,
      "lower": 0.408333,
      "regressed_rows": 0,
      "upper": 0.583333
    },
    "lora_vs_zeroshot_action_accuracy": {
      "delta": 0.758333,
      "improved_rows": 91,
      "lower": 0.683333,
      "regressed_rows": 0,
      "upper": 0.833333
    }
  },
  "receipt_sha256": "bef683521bf43bdd824d85e8db28f6a81109964ddb2fa61961909f89414d0293",
  "training": {
    "adapter_files_sha256": {
      "README.md": "31ebdaaef3eb2d1a1e0a13f5b45ebd26ec50af5d7ef666ba40fc4dc86e20189c",
      "adapter_config.json": "6615c7239d49e28c4e8cd76f86e25c2926923f90707d98bd03d63ba5e17e821c",
      "adapter_model.safetensors": "6bd9dd0792025bafa61f5964a7dbc7a1fc8eed59892a091b12f9bffbc939579b"
    },
    "epochs": 5,
    "first_loss": 1.092918,
    "frozen_base_unchanged": true,
    "last_loss": 6.2e-05,
    "learning_rate": 0.0003,
    "lora_alpha": 32,
    "lora_dropout": 0.05,
    "optimizer_steps": 250,
    "peak_allocated_bytes": 9246205440,
    "peak_reserved_bytes": 11916017664,
    "rank": 16,
    "reload_determinism": {
      "exact_matches": 8,
      "max_abs_logit_diff": 0.0,
      "probes": 8
    },
    "trainable_parameters": 11272192,
    "training_seconds": 315.785
  }
}
