{
  "adapter": {
    "causal_master": {
      "bytes": 931170120,
      "dtypes": [
        "torch.float32"
      ],
      "sha256": "277b27e43ac90a893a03e789232a154d1bc7f84c057eb85ec9cc4e5eb60d7cdf",
      "tensor_count": 256
    },
    "serving_copy": {
      "bytes": 931173960,
      "dtypes": [
        "torch.float32"
      ],
      "sha256": "32aabf31a87935a156676807cc132fb99879d20f66d89b6b39148d5176cb25af",
      "tensor_count": 256
    }
  },
  "base_model": {
    "id": "Qwen/Qwen3.5-9B",
    "revision": "c202236235762e1c871ad0ccb60c8ee5ba337b9a"
  },
  "evaluation": {
    "benchmark": "six-source strict-OOD action-only tool-trajectory suite",
    "mean_ood_auroc": 0.9215356108161963,
    "mean_ood_pauroc_at_20": 0.8211832158363769,
    "pooled_auroc": 0.9355070171618869,
    "pooled_pauroc_at_20": 0.7836630889831644,
    "primary_metric": "unweighted six-source Mean-OOD raw normalized pAUROC@20",
    "rows": 6395,
    "sources": [
      "AgentDojo",
      "CUA-SHADE",
      "IaC",
      "MLE-Sabotage",
      "Rogue Eval",
      "SHADE-Arena"
    ]
  },
  "source_commit": "f408b2ae5c45b743b342e45a0d63a952b30debea",
  "training": {
    "data_rows": 21837,
    "deception_rows": 13149,
    "effective_batch_size": 32,
    "epochs": 1,
    "learning_rate": 5e-05,
    "lora_alpha": 256,
    "max_length": 29696,
    "objective": "binary soft BCE at selected literal 0|1 logits",
    "peak_cuda_memory_allocated_bytes": 44427638784,
    "quantization": "NF4 double-quantized QLoRA with BF16 compute",
    "rank": 128,
    "runtime_seconds": 23946.5855,
    "seed": 0,
    "teacher": "moonshotai/kimi-k3",
    "tool_trajectory_rows": 8688,
    "trainable_parameters": 232783872
  }
}