{
  "schema_version": "sensebench-run-v2",
  "run_id": "vllm-gemma-4-26b-a4b-fp8-h100-p002-lexen-v1-20260614",
  "created_at": "2026-06-14T12:57:33.409617+00:00",
  "git_commit": null,
  "runner": {
    "github_handle": "vassiliphilippov",
    "name": "Vassili Philippov",
    "contact": null
  },
  "dataset": {
    "dataset_id": "lexen",
    "dataset_version": "lexen-v1",
    "dataset_revision": null,
    "content_hash": "sha256:5fd4382b93f19087a1e31f6dd7d1db17c1eb17ff80fcbe1d3fdd55c0c3ecefe8",
    "item_count": 4861
  },
  "prompt": {
    "id": "p002",
    "sensebench_version": "0.1.0"
  },
  "model": {
    "kind": "self_hosted_llm",
    "display_name": "google/gemma-4-26B-A4B-it",
    "requested_model": "hosted_vllm/google/gemma-4-26B-A4B-it",
    "resolved_model": "google/gemma-4-26B-A4B-it",
    "resolved_model_counts": {
      "google/gemma-4-26B-A4B-it": 4871
    },
    "llm_vendor": "Google",
    "source_kind": "open_source",
    "license": "gemma",
    "model_url": "https://huggingface.co/google/gemma-4-26B-A4B-it",
    "reasoning_effort": "no reasoning",
    "hf_revision": "8edbb9269ec9c3faad538ee1208a07eb46051f34",
    "quantization": "fp8",
    "inference_engine": "vllm",
    "inference_engine_version": "0.22.1",
    "container_image": "vllm/vllm-openai:v0.22.1",
    "serve_command": "exec vllm serve RedHatAI/gemma-4-26B-A4B-it-FP8-Dynamic --port 8000 --served-model-name google/gemma-4-26B-A4B-it --max-model-len 8192 --gpu-memory-utilization 0.90 --max-num-seqs 256 --no-enable-prefix-caching --language-model-only --revision 8edbb9269ec9c3faad538ee1208a07eb46051f34",
    "endpoint_base_url": "http://localhost:8000/v1"
  },
  "sampling": {
    "temperature": 0.0,
    "top_p": null,
    "max_tokens": 128,
    "seed": null,
    "extra": {}
  },
  "policy": {
    "votes_per_item": 1,
    "semantic_reasks_per_invalid_vote": 1,
    "tie_break": "earliest_vote",
    "monosemous_policy": "short_circuit"
  },
  "machine": {
    "platform": "Linux-6.17.0-1017-aws-x86_64-with-glibc2.35",
    "cpu_model": "AMD EPYC 7R13 Processor",
    "cpu_cores": 192,
    "ram_gib": 1999.9,
    "gpu": {
      "name": "NVIDIA H100 80GB HBM3",
      "count": 1,
      "vram_mib_per_gpu": 81559,
      "driver_version": "595.71.05",
      "cuda_version": "13.2"
    },
    "provider": "vast.ai",
    "instance_id": "40919268",
    "hourly_rate_usd": 2.2996296296296297
  },
  "execution": {
    "concurrency": 128,
    "warmup_call_count": 8,
    "timing": {
      "benchmark_started_at": "2026-06-14T12:57:13.364920+00:00",
      "benchmark_ended_at": "2026-06-14T12:57:33.407321+00:00",
      "benchmark_seconds": 20.04239618196152,
      "setup_seconds": 2.2232830289867707
    }
  },
  "totals": {
    "item_count": 4861,
    "correct_count": 4241,
    "accuracy": 0.8724542275252005,
    "call_count": 4871,
    "usage": {
      "input_tokens": 1185755,
      "cached_input_tokens": null,
      "output_tokens": 10770,
      "reasoning_output_tokens": null
    },
    "cost": {
      "currency": "USD",
      "total_usd": 0.012802802252448465,
      "input_uncached_usd": null,
      "input_cached_usd": null,
      "output_usd": null,
      "input_uncached_unit_price_usd": null,
      "input_cached_unit_price_usd": null,
      "output_unit_price_usd": null,
      "source": "machine_time_estimate"
    },
    "elapsed_seconds": 20.04239618196152
  }
}