{
  "dataset": "MovieLens latest-small (September 2018 snapshot)",
  "ratings": 100836,
  "users": 610,
  "source_movies": 9742,
  "catalog_movies": 1859,
  "train_ratings": 80896,
  "heldout_ratings": 19940,
  "evaluated_users": 573,
  "relevant_test_items": 6489,
  "model_recall_at_10": 0.1103,
  "popularity_recall_at_10": 0.06097,
  "model_ndcg_at_10": 0.12488,
  "popularity_ndcg_at_10": 0.08299,
  "fit_seconds": 0.44,
  "split": "Latest 20% per user by rating timestamp; ties broken by movie ID. Item eligibility and popularity computed only from training data. Positive = rating >=4. At least 3 training positives and 1 eligible test positive.",
  "evaluation": "All eligible unseen items; no sampled negatives. Evaluate sparse positive top-80 model with the complete training-positive profile. This does not evaluate new users with 1\u20135 picks, group fairness, genre filters or shuffle.",
  "source_sha256": {
    "ratings.csv": "80da8b3393dae325bbba5a31f291a6ba55d8d4f4396de3c456f2c1635b1b70e8",
    "movies.csv": "c47b783f23eb304413307e1bb05335e427737da20ec8d3c0e96335567dd5f3dc",
    "links.csv": "028a4a1d17751ca9276e0e0718cfc22356350f224a996e1ec348c23274b9209d"
  },
  "limitations": [
    "Historical catalog through 2018, skewed toward English-language films and MovieLens participants.",
    "Per-user temporal split is not a single global time cutoff; cross-user information can come from later timestamps.",
    "Offline ratings metrics do not establish real-world satisfaction or reduced decision time.",
    "This is a development snapshot experiment, not a standardized benchmark or state-of-the-art claim."
  ],
  "five_like_evaluation": {
    "users": 563,
    "model_recall_at_10": 0.09822,
    "quality_baseline_recall_at_10": 0.04845,
    "protocol": "Use last five eligible training likes per user, production 0.9 normalized affinity + 0.1 quality ranking; exclude all training-seen movies. Same held-out latest-20% split. Does not represent real new-user satisfaction."
  }
}