Hugging Face's logo

Datasets:
DescribeEvents
/
nba_quarter_events

Tasks:
Time Series Forecasting
Modalities:
Tabular
Text
Formats:
parquet
Languages:
English
Size:
1K - 10K
Tags:
temporal-point-process
event-sequences
nba
sports
play-by-play
marked-temporal-point-process
Libraries:
Datasets
pandas
Polars
License:
Dataset card Data Studio Files
xet
 Community
1
nba_quarter_events / README.md
XiaoBB's picture
XiaoBB
Upload README.md with huggingface_hub
607eafa verified
preview code
|
Raw
History Blame Contribute Delete
10.2 kB
metadata 
language:
  - en
license: mit
task_categories:
  - time-series-forecasting
tags:
  - temporal-point-process
  - event-sequences
  - nba
  - sports
  - play-by-play
  - marked-temporal-point-process
size_categories:
  - 1K<n<10K
pretty_name: NBA Quarter Events
dataset_description: >-
  Curated per-quarter NBA play-by-play sequences from the 2024-25 regular
  season, designed for temporal point process (TPP) and marked temporal point
  process (MTPP) modeling. Each sequence represents one quarter (Q1-Q4) of a
  single game, with the prediction target being the play type of the next event.
  Sequences are filtered to keep only quarters with 90-110 events, ensuring
  consistent sequence lengths. Rich narrative type_text descriptions provide
  fully descriptive context — shot style and distance, make/miss results,
  offensive/defensive rebound classification, full team names on every event,
  and timeout types — enabling an LLM to predict the next event type from the
  sequence alone. Quarter and time information are captured in metadata and
  time_since_start fields rather than repeated in type_text, reducing token
  waste.

NBA Quarter Events

Dataset Description

Curated per-quarter NBA play-by-play sequences from the 2024-25 regular season, designed for temporal point process (TPP) and marked temporal point process (MTPP) modeling. Each sequence represents one quarter (Q1–Q4) of a single game, with the prediction target being the play type of the next event. Sequences are filtered to keep only quarters with 90–110 events, ensuring consistent sequence lengths.

Rich narrative type_text descriptions provide fully descriptive context for each event — shot style and distance, free throw make/miss with sequence position, offensive/defensive rebound classification, team identification on every play, and timeout types — enabling an LLM to predict the next event type from the sequence alone. Quarter and time-remaining information is not repeated in type_text since it is already captured in metadata.quarter and time_since_start.

  • Source: NBA Stats API
  • Season: 2024–25 Regular Season
  • Grouping: Events grouped by quarter (Q1–Q4) within each game
  • Sequences: 2,868
  • Train / Validation / Test: 2,296 / 286 / 286
  • Sequence length: 90–110 events per quarter (mean: 101.5)
  • Total events: 291,091
  • Event types: 9 play categories
  • Time unit: minutes (from quarter start)
  • Source games: 1,225

Dataset Summary

Property Value
Total sequences 2,868
Train / Validation / Test 2,296 / 286 / 286
Event types 9
Total events 291,091
Events per quarter 90–110 (mean 101.5)
Time unit minutes (from quarter start)
Duration per quarter 11.2–12.0 minutes
Season 2024-25 Regular Season
Source games 1,225

Schema

Each record follows the canonical 8-field TPP schema:

Field Type Description
seq_idx int Sequence index (per split)
seq_len int Number of events in the sequence
description str Game summary (teams, date, quarter, player rosters)
metadata str (JSON) game_id, home_team, away_team, home_players, away_players, game_date, season, season_type, time_unit, quarter
time_since_start list[float] Elapsed minutes from the first event in the quarter
time_since_last_event list[float] Minutes since the previous event (first is 0.0)
type_event list[str] Play type label — the prediction target
type_text list[str] Rich narrative description of each event

Event Types (9 categories)

type_event Description Share
rebound Offensive and defensive rebounds 24.3%
missed_fg Missed field goals (2pt or 3pt) 22.4%
made_2pt Made 2-point field goals 13.8%
free_throw Free throw attempts (make/miss) 8.9%
foul Personal, shooting, and offensive fouls 8.3%
made_3pt Made 3-point field goals 7.0%
turnover Turnovers (bad pass, travel, shot clock, etc.) 6.7%
steal_or_block Steals and blocked shots 5.9%
timeout Team timeouts (regular, mandatory) 2.6%

Administrative events (substitution, jump ball, period start/end, violation, ejection, instant replay) are excluded.

type_text (Rich Narrative)

Each event's type_text is a natural language sentence combining multiple context signals. The description is designed to be fully informative so an LLM reading a sequence of events can predict the next event type:

  • Player identity: Name, team full name, position (guard / forward / center)
  • Team identification: Full team name on every event (e.g., "for the Orlando Magic") for unambiguous possession tracking
  • Shot style: Specific shot type from API subType (driving layup, fadeaway jump, pullup jump, cutting dunk, etc.)
  • Shot distance & zone: Distance in feet from the API shotDistance field, plus court zone (paint, left wing, corner three, top of arc, mid-range, etc.)
  • Free throw detail: Make/miss result parsed from description, plus sequence position (e.g., "1 of 2", "2 of 3")
  • Rebound classification: Offensive or defensive, classified via subType and description parsing (covers ~95%+ of rebounds)
  • Timeout team & type: Team identified via location field (home/visitor), with timeout subtype (regular, mandatory)
  • Foul subtype: Lowercased foul type (shooting, personal, offensive, loose ball, etc.)
  • Score context: Leading team, margin, game state (tight game, comfortable margin, blowout)
  • Momentum: Scoring runs with full team name (e.g., "Orlando Magic currently on a 6-0 scoring run"), pattern signals (e.g., "shooting 3 for 5 recently")

Examples:

# Made 2pt — includes shot style, distance, zone, team name
"J. Robinson-Earl. makes a 2-point putback layup from 3 feet near the basket
 for the New Orleans Pelicans. Orlando Magic lead 12-6. early in the game."

# Made 3pt
"K. Ware, Heat center. makes a 3-point jump from 28 feet from the right wing
 for the Miami Heat. Milwaukee Bucks lead 81-65 with a comfortable margin."

# Missed field goal — includes 2pt/3pt distinction and shot style
"J. Robinson-Earl. misses a 2-point layup from 6 feet in the paint for the
 New Orleans Pelicans. The game is tied. early in the game."

# Free throw miss — includes make/miss and sequence position (1 of 2)
"J. Alvarado, Knicks guard. misses free throw 1 of 2 for the New Orleans
 Pelicans. The game is tied."

# Rebound — classified as offensive/defensive
"G. Bitadze, Magic center-forward. grabs an offensive rebound for the Orlando
 Magic. The game is tied. early in the game."

# Turnover — subtype integrated naturally into description
"B. Ingram, Raptors forward. commits a bad pass turnover for the New Orleans
 Pelicans. The game is tied. early in the game."

# Foul — lowercased subtype
"F. Wagner, Magic forward. commits a personal foul for the Orlando Magic.
 The game is tied. early in the game."

# Steal
"K. Caldwell-Pope, Grizzlies guard. records a steal for the Orlando Magic.
 The game is tied. early in the game."

# Timeout — team identified with full name, includes timeout type
"the New Orleans Pelicans call a regular timeout. The game is tied. early in
 the game. Orlando Magic currently on a 6-0 scoring run."

Curation Filters

Quarters are selected from complete games (all 4 regulation quarters present) with diverse, non-trivial event distributions:

Filter Value Description
min-events 90 Min events per quarter
max-events 110 Max events per quarter
min-types 4 At least 4 distinct event types of 9
max-dominant-ratio 0.40 No single event type exceeds 40% of events
max-repeat-ratio 0.30 Consecutive same-type rate below 30%
Completeness — All 4 quarters (Q1–Q4) must be present; overtime excluded

Example 

{
  "seq_idx": 0,
  "seq_len": 96,
  "description": "NBA Regular Season game between the New York Knicks and the Brooklyn Nets on 2025-04-13. Q1 play-by-play events. New York Knicks: C. Payne, L. Shamet, M. Bridges, ...; Brooklyn Nets: D. Timme, K. Johnson, T. Etienne, ...",
  "metadata": "{\"game_id\": \"0022401188\", \"home_team\": \"BKN\", \"away_team\": \"NYK\", \"home_players\": [\"D. Timme\", ...], \"away_players\": [\"C. Payne\", ...], \"game_date\": \"2025-04-13\", \"season\": \"2024-25\", \"season_type\": \"Regular Season\", \"time_unit\": \"minutes\", \"quarter\": \"Q1\"}",
  "time_since_start": [0.0, 0.3167, 0.5333, ...],
  "time_since_last_event": [0.0, 0.3167, 0.2167, ...],
  "type_event": ["foul", "turnover", "made_2pt", ...],
  "type_text": [
    "M. Bridges, Knicks guard-forward. commits a personal take foul for the New York Knicks. The game is tied. early in the game.",
    "commits a shot clock turnover. The game is tied. early in the game.",
    "C. Payne, 76ers guard. makes a 2-point driving finger roll layup from 5 feet in the paint for the New York Knicks. New York Knicks lead 2-0 in a tight game. early in the game.",
    ...
  ]
}

Data Source

Play-by-play data sourced from the NBA Stats API via the nba_api Python package using the PlayByPlayV3 endpoint. Player metadata (name, position, team) fetched via CommonPlayerInfo.

Usage 

from datasets import load_dataset

ds = load_dataset("DescribeEvents/nba_quarter_events")

# Access splits
train = ds["train"]       # 2,296 sequences
val = ds["validation"]    # 286 sequences
test = ds["test"]         # 286 sequences

print(train[0]["type_event"][:5])   # First 5 play types
print(train[0]["type_text"][:5])    # First 5 narrative descriptions

Citation

If you use this dataset, please cite:

@dataset{nba_quarter_events_2025,
  title={NBA Quarter Events: A Temporal Point Process Dataset},
  author={DescribeEvents},
  year={2025},
  publisher={Hugging Face},
  url={https://huggingface.co/datasets/DescribeEvents/nba_quarter_events}
}