luxury-lakehouse/xg-shot-freeze-frames

Pre-Shot xG v3 — Shot Freeze Frames (Context Corpus)

The context half of the training corpus for xg_model_v3, the canonical-SPADL-native pre-shot expected goals model from the luxury-lakehouse analytics platform. One row per (shot, player) — every player present in the shot's freeze frame is a row. The provider is the data_source column, not a separate file format. Sourced from bronze.shot_freeze_frames.

Each shot's freeze-frame player set is joinable to its tabular shot record (dataset xg-shot-data-v3) on the shot identity (match_key, action_id) — action_id is per-match, NOT globally unique, so both keys are always required. The set encoder in xg_model_v3 sum-aggregates this player set into the shot's context vector.

Columns / contract

Column	Type	Meaning
match_key	BIGINT	Kimball match surrogate — half of the shot identity
action_id	BIGINT	Per-match SPADL action id — the other half of the shot identity
data_source	STRING	Provider (statsbomb, skillcorner, gradientsports, ...)
player_id	STRING	Player present in this shot's freeze frame
x	DOUBLE	Player x, canonical SPADL 105×68, home-LTR (goal at x=105)
y	DOUBLE	Player y, canonical SPADL 105×68
is_keeper	INT	1 if this player is a goalkeeper, else 0
is_teammate	INT	1 if this player is on the shooting team, else 0
set_cardinality	INT	Number of players in this shot's freeze-frame set
shooter_attacks_high_x	BOOLEAN	Whether the shooting team attacks the HIGH-x goal in the canonical home-LTR frame (per-shot orientation; may be NULL when it could not be derived)
team_attacking_direction	STRING	Provenance string the shooter_attacks_high_x flag is derived from

Coordinates are canonical SPADL 105×68, home-LTR — no provider is bent to StatsBomb units. StatsBomb-360 freeze frames (raw 120×80) are converted at compute time. One row per (shot, player); the ball row is dropped and the shooter is always included (the sum-aggregation requires actor-inclusion consistency across sources). (access_tier is used internally for the public/restricted split and is dropped before upload.)

Public / restricted split

RM SkillCorner and GradientSports partitions are license-restricted: they publish to a private org-members-only companion repo (xg-shot-freeze-frames-restricted) rather than this public dataset, per lakehouse ADR-049 / ADR-064. StatsBomb-360 freeze frames are public. The split is per-row (per-match access_tier), so a public-licensed SkillCorner match publishes here while a restricted one goes to the companion. A partition migrates here automatically once its license permits public redistribution. The xg_model_v3 trainer reads BOTH repos.

Quick Start

Every row carries a data_source column. The dataset is split into one config per provider, so you can pull a single provider without downloading the rest:

from datasets import load_dataset

# All public providers at once (config "all" — the default):
ds = load_dataset("luxury-lakehouse/xg-shot-freeze-frames", "all", split="train")
df = ds.to_pandas()
print(df["data_source"].value_counts())

# Just one provider (downloads only that provider's parquet):
sb = load_dataset("luxury-lakehouse/xg-shot-freeze-frames", "statsbomb", split="train").to_pandas()

# Reassemble a single shot's freeze-frame set:
shot = sb[(sb["match_key"] == 12345) & (sb["action_id"] == 678)]

Related artifacts

• xg-shot-data-v3 — the tabular shot record (the other half of the corpus), joined on (match_key, action_id)
• spadl-vaep-action-values — the full action-level value corpus

Citation

@software{luxury_lakehouse,
  title  = {Luxury Lakehouse — Serverless Soccer Analytics Platform},
  url    = {https://github.com/karsten-s-nielsen/luxury-lakehouse}
}

License

CC-BY-NC-4.0 — see repository for details.