{
  "paper": "OneRec",
  "arxiv_id": "2502.18965",
  "dataset": "MovieLens 1M",
  "seed": 42,
  "configuration": {
    "sid_cardinalities": [
      256,
      128,
      64
    ],
    "dimensions": 96,
    "experts": 4,
    "session_items": 3,
    "sft_steps": 240,
    "reward_steps": 120,
    "dpo_steps": 80,
    "dpo_pairs": 96,
    "evaluation_users": 200
  },
  "training": {
    "sid_uniqueness": 0.9079870480302212,
    "sft_final_loss": 5.133344602584839,
    "reward_final_loss": 0.6804039984941482,
    "dpo_initial_loss": 0.1357584513258189,
    "dpo_final_loss": 0.0005362891199183651
  },
  "results": {
    "session_generator_sft": {
      "hit_at_10": 0.02,
      "ndcg_at_10": 0.01565464876785729,
      "head_share_at_10": 0.4372163388804841,
      "valid_session_rate": 1.0
    },
    "onerec_iterative_preference_alignment": {
      "hit_at_10": 0.0,
      "ndcg_at_10": 0.0,
      "head_share_at_10": 0.10469314079422383,
      "valid_session_rate": 1.0
    }
  },
  "conclusion": "The complete scaled pipeline ran successfully, but DPO collapsed held-out ranking accuracy despite reducing head concentration and its own objective. The weak local reward model was over-optimized.",
  "schema_version": 2,
  "manifest_ref": "reproduction:onerec",
  "evaluation_protocol": {
    "tier": "l3_paper_pipeline",
    "seeds": [
      42
    ],
    "formal_comparison": false,
    "claim_policy": "single/few-seed smoke result; do not claim a stable improvement"
  },
  "provenance": {
    "artifact_path": "docs/reproductions/2502.18965-onerec/metrics/movielens-1m-seed42.json",
    "historical_migration": "historical-metrics-v2-2026-08-09",
    "original_code_commit": "not recorded",
    "dataset_fingerprint": "not recorded in historical artifact"
  }
}