SportsD: On-Ball Soccer Decision Benchmark

SportsD tests whether a vision-language model picks better on-ball soccer actions than professional players did, using possession-value (VAEP) ground truth from two World Cups.

Each event freezes a real on-ball moment. The model sees the seconds before the moment and picks one action: a pass to a specific lettered teammate, or a shot. Every candidate action has a VAEP expected value, so every answer gets a real value score.

• 1415 events: 726 from the 2022 men's World Cup, 695 from the 2023 women's World Cup.
• Ground truth: the optimal action a* and its value V_optimal, the player's real choice and its value V_human, and the value of every candidate action.
• Stimuli: an anonymized action map, a HUD-masked broadcast clip, and the clip's masked frames per event. Player identities are removed so the model cannot retrieve a known outcome.
• Model responses: per-action belief, value, and expected-value probes for every model.

Layout

answer_key/
  men_ground_truth.csv     one row per event: a*, human choice, V_optimal, V_human, Delta, ...
  men_options.csv          one row per candidate action: EV, type, is_a_star
  women_ground_truth.csv
  women_options.csv
data/
  men/<uid>/anon_map.png   lettered action map, identity removed
  men/<uid>/clip.mp4       HUD-masked broadcast clip
  men/<uid>/frames/*.jpg   masked frames sampled from the clip
  women/<uid>/...
responses/
  men/responses_<model>.json      per-action q1_prob / q2_value / q3_ev for each model
  women/responses_<model>.json
manifest.csv               uid -> split, map path, clip path, frame count

uid is <GAME>__<stem>, for example ARG_CRO_GAME__024m35_Argentina_Fernandez. The same uid keys manifest.csv, answer_key/*, and every entry in responses/*.

Answer key columns

ground_truth.csv: uid, game, stratum, stem, minute, event_uuid, astar_type, astar_letter, human_took, human_letter, V_optimal, V_human, Delta, sd, margin, n_pass_options, actor_x, actor_y, attack_dir, mirrored.

options.csv: uid, game, stratum, action, type, ev, is_a_star. Join on uid to score any answer as regret = V_optimal - EV(answer).

Events are stratified by Delta = V_optimal - V_human (low_vaep, medium_vaep, high_vaep) so the set spans easy and hard decisions.

Scoring a new model

Score every answer from options.csv alone. For one event, let ev be the map from action to EV over that event's candidate rows. Then:

• V_optimal = max(ev) the best action's value
• V_chance = mean(ev) a random pick's value
• EV(answer) = ev[answer] the model's chosen action value
• Regret = V_optimal - EV(answer). Lower is better. Every answer is scoreable.
• Skill = (EV(answer) - V_chance) / (V_optimal - V_chance). 0 = random, 1 = optimal.
• Accuracy = fraction of events where the answer is the argmax of ev (regret = 0).

import pandas as pd, collections
opt = pd.read_csv("answer_key/men_options.csv")
ev  = {u: dict(zip(g.action, g.ev)) for u, g in opt.groupby("uid")}   # uid -> {action: EV}
def score(uid, answer):
    e = ev[uid]; vopt = max(e.values()); vch = sum(e.values())/len(e)
    return dict(regret=vopt - e[answer],
                skill=(e[answer] - vch) / (vopt - vch) if vopt > vch else 1.0,
                correct=e[answer] == vopt)

ground_truth.csv is separate. Its V_optimal, V_human, and Delta are realized-outcome annotations used for stratification and human comparison. For the events where the player already played the optimal action (Delta = 0), ground_truth.V_optimal is the player's realized value, which sits on a different scale than the counterfactual EVs in options.csv. Do not mix the two. Score models from options.csv; read the human's real choice and the strata from ground_truth.csv.

Model responses

Each responses/*/responses_<model>.json maps uid to three ranked probes: q1_prob (pick probability per action), q2_value (value per action), q3_ev (expected value per action). Files with a __text or __lastframe suffix are ablation conditions (text-only context, or a single final frame) rather than the full-clip condition.

License and attribution

Released under CC-BY-4.0. Frames and clips derive from broadcast footage of the 2022 FIFA men's World Cup and the 2023 FIFA women's World Cup; VAEP values come from StatsBomb event data. Cite the SportsD paper when you use this benchmark.