luxury-lakehouse/scoutgpt-training-data

ScoutGPT Training Data — Player Action Sequences

Per-player match-level action sequences unified across StatsBomb Open Data and Wyscout open data. Each row is one player-match's ordered SPADL action sequence with contextual tokens (competition, season, score state, half, opponent), serialized as the token stream that the ScoutGPT transformer consumes during training and at inference.

Part of the (Right! Luxury!) Lakehouse soccer analytics platform.

Quick Start

from datasets import load_dataset

ds = load_dataset("luxury-lakehouse/scoutgpt-training-data")
df = ds["train"].to_pandas()
print(f"{len(df):,} player-match sequences, {df['player_id'].nunique():,} unique players")

Explore interactively: Soccer Analytics App

What Is This Dataset?

ScoutGPT is a transformer decoder trained to produce per-player season embeddings by modelling the sequential structure of on-ball actions within a match. Unlike per-action bag-of-features embeddings (Football2Vec v1), ScoutGPT sees the order of actions — so it can capture tempo, build-up patterns, and decision-making over a possession.

This dataset is the serialized per-player-match training corpus produced by wf-scoutgpt-export from the gold-layer fct_action_values table.

Schema

Column	Type	Description
player_id	Int64	Canonical player identifier (cross-source resolved)
match_id	Int64	Provider-native match identifier
data_source	string	Origin (statsbomb or wyscout)
competition_id	Int64	Competition identifier (NULL for Wyscout)
season_id	Int64	Season identifier (NULL for Wyscout)
team_id	Int64	Player's team in this match
token_ids	list<int32>	Tokenized action-sequence for this player in this match
sequence_length	Int64	Number of tokens in the sequence

Data Sources

Source	Matches	License
StatsBomb Open Data	~3,000	CC-BY 4.0
Wyscout Public Dataset	~1,900	CC-BY-NC 4.0

Inherits the more restrictive CC-BY-NC 4.0 license via Wyscout.

Use Cases

• ScoutGPT training: primary training corpus for the ScoutGPT transformer
• Sequence-aware embedding research: evaluate new architectures (cross-attention, Fourier position encodings, RoPE) against a common corpus
• Downstream fine-tuning: task-specific heads (player-type classification, next-action prediction) on top of pre-trained ScoutGPT checkpoints

Limitations

• Open data only: commercial datasets cover additional leagues and seasons
• Season-level aggregation: per-match sequences are independent — cross-match context is not captured in a single row
• Derived from SPADL: downstream of the SPADL conversion; any SPADL-adapter issue (see spadl-vaep-action-values) propagates here

Companion Resources

Resource	Type	Description
ScoutGPT	Model	Transformer decoder trained on this dataset
ScoutGPT variants (rope)	Model	Ablation checkpoint with RoPE position encoding
SPADL/VAEP Action Values	Dataset	Upstream source — per-action valuations

License

CC-BY-NC 4.0 (inherited from Wyscout).