{
  "methodology": {
    "engine": "vLLM prompt_logprobs=1",
    "dataset": "Salesforce/wikitext",
    "dataset_config": "wikitext-2-raw-v1",
    "split": "test",
    "dataset_fingerprint": "a46124b21ac53738",
    "window_selection": "first contiguous non-empty test-corpus token windows",
    "num_windows": 4,
    "window_tokens": 512,
    "evaluated_next_tokens": 2044,
    "tokenizer": "frozen Gemma 4 QAT/Heretic parent tokenizer",
    "note": "BF16 used vLLM CPU offload on a 24 GB RTX 3090; W4A16 used the same vLLM prompt-logprob scorer without offload."
  },
  "results": {
    "bf16_parent": {
      "perplexity": 388.14988734463424,
      "mean_nll": 5.961391572614971,
      "nll_sum": 12185.084374425001,
      "cpu_offload_gb": 30.0
    },
    "w4a16_g64": {
      "perplexity": 432.31795682768376,
      "mean_nll": 6.069161328695908,
      "nll_sum": 12405.365755854436,
      "cpu_offload_gb": 0.0
    },
    "delta_perplexity": 44.16806948304952,
    "delta_percent": 11.379127219437565
  }
}