Downloads

Event-pipeline JSON (VAEP / JOI / JDI)

Used by Overview, Top Pairs, Team Maps, and the Team Builder.

• data/overview.json tournament-wide stats & top pairs summary Download
• data/pairs.json every qualifying pair with JOI / JDI / minutes Download
• data/teams.json per-team summary, colors, pair counts Download
• data/team_figures_index.json pointers to per-team PNG figures Download
• data/team_builder.json MIP-optimal XI and formation per squad Download
• data/matches.json match-level metadata for the 64-match corpus Download
• data/player_form.json per-player Δ OI/90 vs positional prior Download
• data/chemistry_stories.json narrative blurbs for top teams (one-liners) Download

Attention chemistry (tracking transformer)

Output of the frame-VAEP transformer described in Methodology §2–3. The per-pair and per-group attention totals are what the Chemistry Leaderboards render.

• data/attention_pairs.json pair attention chemistry — total, per-90, lift vs team baseline Download
• data/attention_groups.json triples + quads with attention lift Download
• data/attention_figures_index.json per-team attention heatmap PNG pointers Download
• data/chemistry_concepts.json 12-mechanism tactical taxonomy (third-man triangle, overlap, etc.) Download
• data/vaep_metrics_transformer.json frame-VAEP head val metrics (AUC / Brier / loss) Download

Club vs National & FIFA Mode

Cross-context comparisons (StatsBomb open data passed through the same VAEP model) and the FIFA-rating crosswalk used by FIFA Mode.

• data/cross_context.json per-player OI/90 across WC22 / Euro 2024 / Copa 2024 / club Download
• data/cross_chem.json cross-context chemistry deltas Download
• data/fifa_mode.json FIFA 23 vs FIFA 26 national-team ratings crosswalk Download

data/wc26_rosters.json ships when the FIFA-26 roster rebuild lands — link will appear here automatically.

Digital Whiteboard

Counterfactual swap outputs from the frame-VAEP transformer — what happens to predicted scoring probability when you sub one player out for another over a real possession sequence.

• data/whiteboard_moves.json precomputed counterfactual swap deltas Download

Tabular data (CSV)

Generated from parquet sources by the export pipeline.

• CSV manifest will appear here when data/downloads.json is present.

Figures (PNG)

Per-team offensive and defensive chemistry maps.

• PNG manifest will appear here when data/team_figures_index.json is present.

Model artifacts

Event-pipeline classifiers (joblib) plus PyTorch Lightning checkpoints for the tracking transformer. Loading recipes are in Methodology §6.

• vaep_bundle.joblib trained score & concede classifiers + feature spec (event VAEP) Download
• predictor.joblib expected-OI-per-zone regressor used for JDI Download
• training_metrics_xt.json xT-regression val metrics (Spearman ρ = 0.714, MAE, lift over xT-lookup) Download
• training_metrics_frame_vaep.json frame-VAEP val metrics (BCE AUC ≈ 0.80 / 0.79 for score / concede heads) Download

Transformer checkpoints (transformer_xt_regression.ckpt, transformer_frame_vaep.ckpt) are gitignored to keep the repo light — rebuild them from the repro recipe in ~10 minutes on a single GPU, or contact us for a direct link.