benchmarking/tinystories_10m_eval_metrics.json

{
  "model": "10M",
  "num_samples": 100,
  "temperature": 0.8,
  "top_k": 50,
  "metrics": {
    "ttft_ms": 5.45,
    "tps": 296.9,
    "total_latency_ms": 90.2,
    "tokens_generated": 26.7,
    "perplexity": 4.13,
    "avg_token_prob": 0.4957,
    "repetition_rate": 0.1298,
    "coherence_length": 21.7,
    "vocab_diversity": 0.8235
  },
  "evaluated_at": "2026-06-21 09:13:07"
}