goalmind/penalty-kick-dataset

⚽ GolaMind — Penalty Kick Dataset

Overview

Synthetic dataset of 10,100 penalty kicks built for the GolaMind AI application — an advanced tactical analysis tool that helps goalkeepers predict where a penalty taker will shoot.

The dataset was generated using a pre-trained Large Language Model (Qwen2.5-0.5B-Instruct) with three distinct prompting techniques (Zero-Shot, Few-Shot, and Chain-of-Thought), combined with deterministic validation logic to ensure statistical realism.

---

Dataset Statistics

Metric	Value
Total kicks	10,100
Unique players	201
Kicks per player	50
Leagues covered	3
Features per kick	22
Seasons covered	2021/22 – 2024/25

---

Leagues & Teams

League	Players	Teams
UEFA Champions League (UCL)	62	Real Madrid, PSG, Arsenal, Bayern Munich, Man City, Inter Milan, Barcelona, Atletico Madrid, Dortmund, Porto
Premier League	75	Arsenal, Man City, Man United, Liverpool, Chelsea, Tottenham, Aston Villa, Newcastle, Brighton, Brentford, Fulham, Bournemouth, Crystal Palace, Nottm Forest, West Ham, Sunderland
La Liga	64	Barcelona, Real Madrid, Atletico Madrid, Villarreal, Real Betis, Real Sociedad, Athletic Bilbao, Celta Vigo, Sevilla, Osasuna, Getafe, Rayo Vallecano, Valencia, Espanyol

---

Features

Player Profile Features

Feature	Description	Values
player_name	Full name of the player	string
team	Club team	string
league	Competition	UCL / Premier League / La Liga
nationality	Player nationality	string
dominant_foot	Preferred kicking foot	right / left
kick_style	Style of penalty kick	power / placement / chip / panenka
keeper_dependency	Does player react to keeper movement?	independent / dependent
pressure_behavior	How player performs under pressure	consistent / shaky / clutch
reputation_score	Player quality rating	6–10

Kick Features

Feature	Description	Values
season	Football season	2021/22 – 2024/25
kick_number	Kick number in player history	1–50
kick_in_sequence	Position in sequence	1–50
match_stage	Stage of the match	regular / late_close / shootout / final
minute	Match minute	0–90
pressure_level	Pressure intensity	low / medium / high / very_high
zone	Target zone in goal	left_low / left_high / center_low / center_high / right_low / right_high
direction	Horizontal direction	left / center / right
height	Vertical height	low / high
kick_speed_kmh	Estimated kick speed	30–120 km/h
run_up_steps	Number of run-up steps	3–8
prev_kick_zone	Zone of previous kick	zone or "none"

Outcome Features

Feature	Description	Values
outcome	Result of the kick	goal / saved / miss
is_goal	Boolean goal flag	True / False
was_saved	Boolean saved flag	True / False

---

Goal Zone Map

GOALKEEPER'S PERSPECTIVE

+------------------+------------------+------------------+
|                  |                  |                  |
|    LEFT HIGH     |   CENTER HIGH    |   RIGHT HIGH     |
|                  |                  |                  |
+------------------+------------------+------------------+
|                  |                  |                  |
|    LEFT LOW      |   CENTER LOW     |   RIGHT LOW      |
|                  |                  |                  |
+------------------+------------------+------------------+

---

Synthetic Data Generation Method

This dataset was generated using a Large Language Model (LLM) pipeline with three distinct prompting techniques:

Model Used

Qwen/Qwen2.5-0.5B-Instruct — a state-of-the-art 0.5B-parameter instruction-following model chosen for its excellent JSON output quality and ability to run in ~1GB VRAM on a Colab T4 GPU.

Prompting Techniques

1. Zero-Shot Prompting — Player Profile Generation

Used to generate realistic player profiles (name, nationality, foot, style, pressure behavior). Why Zero-Shot? Player profiles require creativity and diversity — providing examples would bias the output toward specific patterns.

2. Few-Shot Prompting — Standard Player Kick Records

Used for players with reputation score 6–7 (consistent/neutral pressure behavior). Why Few-Shot? Examples enforce the exact JSON schema and realistic distributions.

Three carefully chosen example kicks prime the model on format and style before generating new records.

3. Chain-of-Thought (CoT) Prompting — Star & Complex Players

Used for players with reputation ≥ 8, or pressure_behavior in {clutch, shaky}. Why CoT? Forces the model to reason step-by-step about player tendencies, producing richer behavioral patterns.

The model reasons inside <think>...</think> tags about zone selection, outcome rates, and stage distribution before outputting the JSON array.

Generation Pipeline

• Phase 1: Zero-Shot → 201 Player Profiles (one LLM call per team)
• Phase 2: Few-Shot / CoT → 50 Kick Records per Player (5 batches × 10 kicks)
  • Reputation ≥ 8 or clutch/shaky → Chain-of-Thought
  • Otherwise → Few-Shot
• Phase 3: Profile-Guided Fallback → Statistical simulation for any failed LLM calls
• Phase 4: Validation → JSON parsing, schema enforcement, statistical bounds checking

---

Key EDA Findings

Outcome Distribution

Zone Distribution

Goal Rate by Kick Style & Pressure

Top 10 Players by Goal Rate

Keeper Dependency & Pressure Behavior

Zone Heatmap by League & Match Stage

Sequential Bias Analysis

Key Numbers

• Goal rate: 75.3% — consistent with real-world penalty statistics
• Save rate: 21.4% | Miss rate: 3.3%

Zone Distribution

• right_low is the most common zone (37%) — most players are right-footed and aim at their natural corner
• right_high: 16% | left_low: 14% | center zones: ~10% each

Goal Rate by Kick Style

Style	Goal Rate
Placement	75.6%
Power	75.2%
Chip	75.1%
Panenka	69.0%

Panenka carries the highest risk — easily saved if the keeper reads it

Keeper Dependency

Type	Goal Rate
Independent	75.5%
Dependent	73.1%

Independent kickers decide direction before the run-up — harder to stop

Pressure Behavior

Type	Goal Rate
Clutch	75.6%
Consistent	75.6%
Shaky	73.1%

Shaky players drift toward center zones under high pressure — easier to save

Zone Repetition

• Players repeat the same zone ~28-29% of the time
• Sequential bias: after kicking to one side, slight tendency to switch

---

Files

File	Description	Size
penalty_kicks.csv	Main dataset — 10,100 kicks	~1.6 MB
player_profiles.json	201 player profiles with full behavioral DNA	~70 KB
embeddings.npy	Sentence embeddings (all-mpnet-base-v2, 768-dim)	~620 KB
player_names.json	Sorted list of all 201 player names	~3.5 KB
Final_Project.ipynb	Full generation + EDA notebook	—
plot_*.png	EDA visualizations (7 plots)	—

---

Live Application

Try the live GolaMind app: https://huggingface.co/spaces/goalmind/penalty-kick-predictor

Enter any player name and receive:

• Zone probability heatmap (goalkeeper perspective)
• Full statistical analysis (50 kicks)
• AI tactical recommendation powered by Claude

---

Citation

@dataset{goalmind2026,
  title   = {GolaMind Penalty Kick Dataset},
  author  = {GolaMind Team},
  year    = {2026},
  url     = {https://huggingface.co/datasets/goalmind/penalty-kick-dataset},
  note    = {Generated with Qwen2.5-0.5B-Instruct using Zero-Shot, Few-Shot, and Chain-of-Thought prompting}
}