Datasetpublic

gilberty005/interactive-sports-nhl

interactive-sports: NHL research database The database the agents in interactive_sports query. One SQLite file, 2.46 GB, covering 2010-10-07 to 2026-06-14. Agents never see this file directly. The harness builds cutoff-scoped, tokenised VIEWS over it — every view is filtered to game_date <= as_of_date, and every player and team is replaced by an opaque P#### / T#### token that is minted fresh per run. The raw tables below carry real identities; the agent surface does not.… See the full description on the dataset page: https://huggingface.co/datasets/gilberty005/interactive-sports-nhl.

sourceHugging Faceotherupdated 18d agoView on Hugging Face

0likes39downloads

Dataset Card

interactive-sports: NHL research database

The database the agents in interactive_sports query. One SQLite file, 2.46 GB, covering 2010-10-07 to 2026-06-14.

Agents never see this file directly. The harness builds cutoff-scoped, tokenised VIEWS over it — every view is filtered to game_date <= as_of_date, and every player and team is replaced by an opaque P#### / T#### token that is minted fresh per run. The raw tables below carry real identities; the agent surface does not.

Verifying a download

python -m data.download --nhl

checks size and SHA-256 against the manifest. A truncated SQLite file still opens — it just answers some queries and not others — so "it downloaded" is not the same as "it downloaded intact".

sha256 cf7eb0a621a6937111300a0daeb4d2ff2ae28fce471904295a4763ce32d10be5 bytes 2,456,731,648

Tables

table	rows
`shift`	15,556,630
`event`	6,478,514
`skater_game`	739,588
`skater_game_adv`	739,552
`note`	126,967
`edge_shot_location`	110,418
`goalie_game`	82,189
`goalie_game_adv`	44,036
`team_game_adv`	41,094
`game`	20,547
`fixture`	20,485
`season_roster`	16,396
`team_standing`	12,581
`edge_skater_season`	4,733
`player`	3,260
`edge_goalie_season`	525
`team`	36

Built by data_generation/build_nhl_db.py + backfill_nhl.py from the public NHL API. Distributed for research reproducibility: re-scraping does not reproduce the same bytes, and different bytes mean different player tokens, which makes results from before and after a rebuild incomparable.