ardge-hf/crag-mock-kg

CRAG Mock Knowledge Graph (Parquet mirror)

Parquet re-packaging of the Mock API / Knowledge Graph backing data from Meta's CRAG: Comprehensive RAG Benchmark (KDD Cup 2024), used by the mock_api FastAPI server for Task 2 and Task 3. This is the live-queryable structured knowledge side of CRAG, complementary to the per-question web snapshots in ardge-hf/crag-qa — a Task 3 answer may need either or both sources; see that repo's card for the distinction.

Each of the 5 domains is a heterogeneous collection of tables/dicts in the upstream repo (31 total: 5 finance sqlite tables, 15 NBA sqlite tables + 1 soccer table, 3 movie JSON dicts, 6 music pickles, 1 Wikipedia-style open KG). To let every domain live as one HF split despite very different original shapes, every row here is normalized to the same 3-column schema:

• entity_id — the natural key of the row (ticker, player/game/team id, movie/person name, artist/song name, Wikipedia entity name; composite keys joined with | where the source had none, e.g. NBA play_by_play uses game_id|eventnum)
• record_type — which original table/dict the row came from (e.g. price_history, play_by_play, person, grammy_award)
• data — the original record, JSON-serialized as a string

Filter by record_type to recover one original table.

Splits

split	rows	record_types
finance	45,485	company_name, price_history, detailed_price_history, dividend_history, market_cap, financial_info
sports	14,062,822	team, draft_combine_stats, team_details, draft_history, team_history, common_player_info, inactive_players, player, line_score, game, other_stats, game_info, game_summary, officials, play_by_play (13,592,899 — full NBA historical play-by-play log), soccer_match
movie	437,926	movie, person, year
music	5,041,729	artist, song, artist_works, billboard_rank_by_position, billboard_history_by_song, grammy_award
open	3,355	entity (Wikipedia-infobox-style summaries)

from datasets import load_dataset

finance = load_dataset("ardge-hf/crag-mock-kg", split="finance")
nba_games = finance.filter(lambda r: False)  # example only; use sports split
sports = load_dataset("ardge-hf/crag-mock-kg", split="sports")
play_by_play = sports.filter(lambda r: r["record_type"] == "play_by_play")

Notes on the source data

• The upstream Mock API also exposes parameterized fuzzy-name lookup (BM25) over some of these tables (e.g. FinanceKG.get_ticker_by_name, OpenKG.search_entity_by_name) — that lookup logic itself is not part of this mirror, only the underlying data it searches over.
• grammy_df.pickle in the upstream repo fails to unpickle under current pandas (BlockManager placement format changed since it was pickled); it was recovered here via a compatibility shim and is included as record_type=grammy_award.
• KG snapshots are not necessarily time-aligned to any specific question's query_time in the QA set — e.g. a market-cap lookup here may not exactly match a QA ground-truth value captured at a different moment.

License

Original data licensed under CC BY-NC 4.0 by Meta Platforms, Inc. Non-commercial use only; retain attribution.

Citation

@article{yang2024crag,
      title={CRAG -- Comprehensive RAG Benchmark},
      author={Xiao Yang and Kai Sun and Hao Xin and Yushi Sun and Nikita Bhalla and Xiangsen Chen and Sajal Choudhary and Rongze Daniel Gui and Ziran Will Jiang and Ziyu Jiang and Lingkun Kong and Brian Moran and Jiaqi Wang and Yifan Ethan Xu and An Yan and Chenyu Yang and Eting Yuan and Hanwen Zha and Nan Tang and Lei Chen and Nicolas Scheffer and Yue Liu and Nirav Shah and Rakesh Wanga and Anuj Kumar and Wen-tau Yih and Xin Luna Dong},
      year={2024},
      journal={arXiv preprint arXiv:2406.04744},
      url={https://arxiv.org/abs/2406.04744}
}