Dataset Viewer
Auto-converted to Parquet Duplicate
bytes
int64
sha256
string
rows
dict
coverage
dict
2,456,731,648
cf7eb0a621a6937111300a0daeb4d2ff2ae28fce471904295a4763ce32d10be5
{ "edge_goalie_season": 525, "edge_shot_location": 110418, "edge_skater_season": 4733, "event": 6478514, "fixture": 20485, "game": 20547, "goalie_game": 82189, "goalie_game_adv": 44036, "note": 126967, "player": 3260, "season_roster": 16396, "shift": 15556630, "skater_game": 739588, "skater_...
{ "first_game": "2010-10-07T00:00:00", "last_game": "2026-06-14T00:00:00" }

interactive-sports: NHL research database

One SQLite file, 2.46 GB, covering 2010-10-07 to 2026-06-14: 16 tables and ~23M rows of NHL box scores, play-by-play, shifts, and 126,967 dated news notes. It is the database the agents in interactive_sports query.

Agents never read it directly. The harness builds cutoff-scoped views over it, filtered to game_date <= as_of_date, with every player and team replaced by an opaque P#### / T#### token minted fresh per run.

Use

python -m data.download

Downloads and checks the file against nhl_manifest.json, which also carries the per-table row counts. Verify rather than trust the transfer: a truncated SQLite file still opens, and answers some queries but not others.

sha256  cf7eb0a621a6937111300a0daeb4d2ff2ae28fce471904295a4763ce32d10be5
bytes   2,456,731,648

Built by data_generation/build_nhl_db.py from the public NHL API. Distributed for reproducibility: re-scraping does not reproduce the same bytes, and different bytes mean different player tokens.

Downloads last month
43