benchmark_results.json

{
  "baseline_ppl": 430.95936849172716,
  "qt_ppl": 6186.4903309265665,
  "baseline_params": 1554570,
  "qt_params": 793882,
  "compression": 1.9581877407473656,
  "entropies": [
    1.317113995552063,
    1.3771847486495972,
    1.364831566810608,
    1.2537084817886353,
    1.262764573097229,
    1.4013395309448242,
    1.2422407865524292,
    1.6306532621383667,
    1.2806912660598755,
    1.344775676727295,
    1.1928328275680542,
    1.6661663055419922,
    1.30397367477417,
    1.3749926090240479,
    1.4987125396728516,
    1.6543607711791992,
    1.367347002029419,
    1.133209228515625,
    1.2703715562820435,
    0.8554855585098267
  ],
  "ranks": [
    3,
    3,
    3,
    3,
    3,
    3,
    3,
    3,
    3,
    3,
    3,
    3,
    3,
    3,
    3,
    3,
    3,
    3,
    3,
    2
  ],
  "blocktt_active": true,
  "quantum_savings": {
    "n_quantum_tokens": 4,
    "n_total_tokens": 20,
    "quantum_ratio": 0.20000000298023224,
    "mean_routing_score": 0.5120605230331421,
    "mean_entropy": 1.3396377563476562
  },
  "base_latency_ms": 6.601548194885254,
  "qt_latency_ms": 13.520391782124836
}