{
  "adapter": {
    "causal_master": {
      "bytes": 2038466576,
      "dtypes": [
        "torch.float32"
      ],
      "sha256": "0716efc123175c2714ebc24a16d5348a4e357050036f99d8568293d74d6ad579",
      "tensor_count": 256
    },
    "serving_copy": {
      "bytes": 2038470416,
      "dtypes": [
        "torch.float32"
      ],
      "sha256": "4eec7b62b9af53f144bf727bb77c204bcaa71c770766e97b360bb4536a1fd775",
      "tensor_count": 256
    }
  },
  "base_model": {
    "id": "Qwen/Qwen3.5-4B",
    "revision": "851bf6e806efd8d0a36b00ddf55e13ccb7b8cd0a"
  },
  "evaluation": {
    "benchmark": "six-source strict-OOD action-only tool-trajectory suite",
    "bf16_serving": {
      "id_macro_auroc": 0.9782143570726712,
      "id_macro_pauroc_at_20": 0.9331950861436367,
      "mean_ood_auroc": 0.9210723661622273,
      "mean_ood_pauroc_at_20": 0.803876420508685,
      "pooled_auroc": 0.9383858013950402,
      "pooled_pauroc_at_20": 0.7670527336884218
    },
    "id_macro_auroc": 0.9710830300280053,
    "id_macro_pauroc_at_20": 0.9116742836759423,
    "id_rows": 3012,
    "mean_ood_auroc": 0.9137947103322152,
    "mean_ood_pauroc_at_20": 0.770274594973614,
    "pooled_auroc": 0.9173121618351605,
    "pooled_pauroc_at_20": 0.7004981300464136,
    "primary_metric": "unweighted six-source Mean-OOD raw normalized pAUROC@20",
    "primary_serving": "optimized (FP8 attention projections, FP4 MLP/GDN, MXFP8 attention)",
    "rows": 6395,
    "source": "results/qad-v02-optimized01 and results/qad-v02-bf16-01",
    "sources": [
      "AgentDojo",
      "CUA-SHADE",
      "IaC",
      "MLE-Sabotage",
      "Rogue Eval",
      "SHADE-Arena"
    ]
  },
  "source_commit": "67c9ba502d8e02761177ad0a255932bdb8c71ca4",
  "training": {
    "data_filter": "strict teacher/source-label agreement on the 8,688 tool-trajectory rows",
    "data_rows": 7836,
    "deception_rows": 0,
    "effective_batch_size": 32,
    "epochs": 1,
    "learning_rate": 5e-05,
    "lora_alpha": 768,
    "max_length": 29696,
    "objective": "binary soft BCE at selected literal 0|1 logits",
    "peak_cuda_memory_allocated_bytes": null,
    "quantization": "QAD through NVFP4 MLP/GDN, FP8 attention projections and MXFP8 attention; FP32 master LoRA",
    "rank": 384,
    "recipe": "v0.2 (exact three-seed average) followed by same-model quantization-aware distillation: 245 updates, KL T=1, AdamW 1e-5, distilling v0.2's own BF16 scores on the 7,836 training rows through the optimized-serving quantization",
    "runtime_seconds": 3859.4864804700046,
    "seed": [
      0,
      1,
      2
    ],
    "teacher": "moonshotai/kimi-k3",
    "tool_trajectory_rows": 7836,
    "trainable_parameters": null,
    "weight_average": "rank-384 LoRA stacking the three rank-128 seed adapters (A stacked, B/3, alpha 768); exactly the mean update, max relative error 2.2e-8"
  }
}