results/html_token_stats.json

{
  "model_id": "Qwen/Qwen3-VL-8B-Instruct",
  "websight": {
    "name": "websight",
    "n": 10000,
    "min": 164,
    "max": 1483,
    "mean": 455.8,
    "p50": 443,
    "p90": 621,
    "p95": 688,
    "p99": 851
  },
  "gt_html": {
    "name": "gt_html (Design2Code eval GT)",
    "n": 484,
    "min": 564,
    "max": 80030,
    "mean": 24115.1,
    "p50": 20868,
    "p90": 52846,
    "p95": 61371,
    "p99": 73392
  },
  "merged": {
    "name": "merged (websight + gt_html)",
    "n": 10484,
    "min": 164,
    "max": 80030,
    "mean": 1548.0,
    "p50": 450,
    "p90": 675,
    "p95": 981,
    "p99": 39917
  },
  "recommend_max_html_tokens": {
    "conservative_cover_p99": 8192,
    "balanced_cover_p95": 1024,
    "note": "train_compressor --max_html_tokens should be >= chosen percentile to avoid losing tail supervision"
  }
}