benchmarking/tinystories_7m_eval_metrics.json

{
  "model": "7M",
  "num_samples": 100,
  "temperature": 0.8,
  "top_k": 50,
  "metrics": {
    "ttft_ms": 7.75,
    "tps": 168.6,
    "total_latency_ms": 178.5,
    "tokens_generated": 30.3,
    "perplexity": 3.81,
    "avg_token_prob": 0.5125,
    "repetition_rate": 0.1223,
    "coherence_length": 23.2,
    "vocab_diversity": 0.8124
  },
  "evaluated_at": "2026-06-21 09:21:41"
}