{
  "date": "2026-11-05",
  "timezone": "Asia/Tokyo",
  "hardware": {
    "cpu": {
      "Name": "Intel(R) Core(TM) i5-14500HX",
      "NumberOfCores": 14,
      "NumberOfLogicalProcessors": 20
    },
    "ramGiB": 31.8,
    "gpu": "NVIDIA GeForce RTX 4070, 12282 MiB, 621.01",
    "runtime": "llama.cpp CUDA12 2.69.0"
  },
  "software": {
    "node": "22.22.1",
    "lmstudioSdk": "2.1.0",
    "runtime": "llama.cpp CUDA12 2.69.0"
  },
  "models": [
    {
      "type": "llm",
      "modelKey": "plamo-3-translate-8b",
      "format": "gguf",
      "displayName": "Plamo 3 Translate 8B",
      "publisher": "pfnet",
      "path": "pfnet/plamo-3-translate-8b-GGUF/plamo-3-translate-8b-Q4_K_M.gguf",
      "sizeBytes": 5280435168,
      "indexedModelIdentifier": "pfnet/plamo-3-translate-8b-GGUF/plamo-3-translate-8b-Q4_K_M.gguf",
      "deviceIdentifier": null,
      "paramsString": "8.5B",
      "architecture": "plamo3",
      "quantization": {
        "name": "Q4_K_M",
        "bits": 4
      },
      "vision": false,
      "trainedForToolUse": false,
      "maxContextLength": 262144
    },
    {
      "type": "llm",
      "modelKey": "plamo-3-translate-2b",
      "format": "gguf",
      "displayName": "Plamo 3 Translate 2B",
      "publisher": "pfnet",
      "path": "pfnet/plamo-3-translate-2b-GGUF/plamo-3-translate-2b-Q4_K_M.gguf",
      "sizeBytes": 1807192480,
      "indexedModelIdentifier": "pfnet/plamo-3-translate-2b-GGUF/plamo-3-translate-2b-Q4_K_M.gguf",
      "deviceIdentifier": null,
      "paramsString": "2.9B",
      "architecture": "plamo3",
      "quantization": {
        "name": "Q4_K_M",
        "bits": 4
      },
      "vision": false,
      "trainedForToolUse": false,
      "maxContextLength": 262144
    },
    {
      "type": "llm",
      "modelKey": "plamo-3-translate-0.8b",
      "format": "gguf",
      "displayName": "Plamo 3 Translate 0.8B",
      "publisher": "pfnet",
      "path": "pfnet/plamo-3-translate-0.8b-GGUF/plamo-3-translate-0.8b-Q4_K_M.gguf",
      "sizeBytes": 573116288,
      "indexedModelIdentifier": "pfnet/plamo-3-translate-0.8b-GGUF/plamo-3-translate-0.8b-Q4_K_M.gguf",
      "deviceIdentifier": null,
      "paramsString": "890M",
      "architecture": "plamo3",
      "quantization": {
        "name": "Q4_K_M",
        "bits": 4
      },
      "vision": false,
      "trainedForToolUse": false,
      "maxContextLength": 262144
    },
    {
      "type": "llm",
      "modelKey": "gemma-4-e2b-it-qat",
      "format": "gguf",
      "displayName": "Gemma 4 E2B Instruct QAT",
      "publisher": "lmstudio-community",
      "path": "lmstudio-community/gemma-4-E2B-it-QAT-GGUF/gemma-4-E2B-it-QAT-Q4_0.gguf",
      "sizeBytes": 4336348704,
      "indexedModelIdentifier": "lmstudio-community/gemma-4-E2B-it-QAT-GGUF/gemma-4-E2B-it-QAT-Q4_0.gguf",
      "deviceIdentifier": null,
      "paramsString": "4.6B",
      "architecture": "gemma4",
      "quantization": {
        "name": "Q4_0",
        "bits": 4
      },
      "vision": true,
      "trainedForToolUse": true,
      "maxContextLength": 131072
    },
    {
      "type": "llm",
      "modelKey": "gemma-4-e4b-it-qat",
      "format": "gguf",
      "displayName": "Gemma 4 E4B Instruct QAT",
      "publisher": "lmstudio-community",
      "path": "lmstudio-community/gemma-4-E4B-it-QAT-GGUF/gemma-4-E4B-it-QAT-Q4_0.gguf",
      "sizeBytes": 6146492960,
      "indexedModelIdentifier": "lmstudio-community/gemma-4-E4B-it-QAT-GGUF/gemma-4-E4B-it-QAT-Q4_0.gguf",
      "deviceIdentifier": null,
      "paramsString": "7.5B",
      "architecture": "gemma4",
      "quantization": {
        "name": "Q4_0",
        "bits": 4
      },
      "vision": true,
      "trainedForToolUse": true,
      "maxContextLength": 131072
    },
    {
      "type": "llm",
      "modelKey": "google/gemma-4-12b-qat",
      "format": "gguf",
      "displayName": "Gemma 4 12B QAT",
      "publisher": "google",
      "path": "google/gemma-4-12b-qat",
      "sizeBytes": 7151066820,
      "indexedModelIdentifier": "google/gemma-4-12b-qat",
      "deviceIdentifier": null,
      "paramsString": "12B",
      "architecture": "gemma4",
      "quantization": {
        "name": "Q4_0",
        "bits": 4
      },
      "variants": [
        "google/gemma-4-12b-qat@q4_0"
      ],
      "selectedVariant": "google/gemma-4-12b-qat@q4_0",
      "vision": true,
      "trainedForToolUse": true,
      "maxContextLength": 262144
    }
  ],
  "protocol": {
    "created": "2026-11-05T08:05:40.739Z",
    "contextLength": 4096,
    "repetitions": 3,
    "temperature": 0,
    "seed": 42,
    "reasoningBudget": false,
    "contextOverflowPolicy": "stopAtLimit",
    "maxTokens": "4096 minus measured prompt tokens minus 32 reserve; same total context for all",
    "cachePolicy": "Before EVERY warmup/measured translation: raw completion of unrelated cache-reset text with 1 output token replaces active slot; no conversation history; context checkpoints disabled.",
    "warmup": "one translation in each direction for each condition; excluded",
    "models": [
      {
        "id": "plamo-0.8b",
        "key": "plamo-3-translate-0.8b"
      },
      {
        "id": "plamo-2b",
        "key": "plamo-3-translate-2b"
      },
      {
        "id": "plamo-8b",
        "key": "plamo-3-translate-8b"
      },
      {
        "id": "gemma-e2b",
        "key": "gemma-4-e2b-it-qat",
        "draft": "[MODEL_ROOT]/amaranus/Gemma-4-E2B-it-qat-assistant-MTP-Q8_0-GGUF/gemma-4-E2B-it-qat-assistant-MTP-Q8_0.gguf"
      },
      {
        "id": "gemma-e4b",
        "key": "gemma-4-e4b-it-qat",
        "draft": "[MODEL_ROOT]/unsloth/gemma-4-E4B-it-GGUF/mtp-gemma-4-E4B-it-Q8_0.gguf"
      },
      {
        "id": "gemma-12b",
        "key": "google/gemma-4-12b-qat",
        "draft": "[MODEL_ROOT]/yuxinlu1/gemma-4-12B-agentic-fable5-composer2.5-v2-3.5x-tau2-GGUF/gemma-4-12B-it-MTP-Q8_0.gguf"
      }
    ],
    "notes": [
      "Synthetic original corpus, 12 texts, 4 short/4 medium/4 long. Not a standardized external translation benchmark.",
      "Quality assessed separately from timing; context/output exhaustion is a failure, never silently truncated input.",
      "All Gemma runs use external assistant head via --spec-type draft-mtp.",
      "All models use one slot, full requested GPU offload, Flash Attention, f16 KV; record actual GPU data and runtime logs.",
      "Model order fixed; text order deterministically rotated/reversed per repetition. Results are within this session, not thermal-order-controlled."
    ]
  },
  "resolvedLoadSettings": {
    "plamo-0.8b": {
      "contextLength": 4096,
      "evalBatchSize": 2048,
      "physicalBatchSize": 512,
      "parallel": 1,
      "flashAttention": true,
      "offloadKVCacheToGpu": true
    },
    "plamo-2b": {
      "contextLength": 4096,
      "evalBatchSize": 2048,
      "physicalBatchSize": 512,
      "parallel": 1,
      "flashAttention": true,
      "offloadKVCacheToGpu": true
    },
    "plamo-8b": {
      "contextLength": 4096,
      "evalBatchSize": 2048,
      "physicalBatchSize": 512,
      "parallel": 1,
      "flashAttention": true,
      "offloadKVCacheToGpu": true
    },
    "gemma-e2b": {
      "contextLength": 4096,
      "evalBatchSize": 2048,
      "physicalBatchSize": 512,
      "parallel": 1,
      "flashAttention": true,
      "offloadKVCacheToGpu": true
    },
    "gemma-e4b": {
      "contextLength": 4096,
      "evalBatchSize": 2048,
      "physicalBatchSize": 512,
      "parallel": 1,
      "flashAttention": true,
      "offloadKVCacheToGpu": true
    },
    "gemma-12b": {
      "contextLength": 4096,
      "evalBatchSize": 2048,
      "physicalBatchSize": 1120,
      "parallel": 1,
      "flashAttention": true,
      "offloadKVCacheToGpu": true
    }
  },
  "notes": [
    "No model weights are included.",
    "Personal filesystem locations are replaced by [MODEL_ROOT]; API credentials and process identifiers are not part of this export.",
    "runs.jsonl includes text emitted by the tested local models as reasoning; it is NOT the benchmark author or grader's private reasoning.",
    "342 timed trials: 324 standard plus 18 extra-long. 126 quality judgments: 14 distinct inputs times 9 conditions.",
    "Raw stop times on truncated runs are not successful translation completion times.",
    "CPU thread pools and physical microbatch settings differ as reported in the article."
  ]
}