choucsan/FIFA_World_Cup_Games

FIFA World Cup Full-Match Video Dataset

This dataset provides YouTube full-match references and structured match annotations for classic FIFA World Cup games. Instead of redistributing video files, it stores YouTube video IDs and URLs, then links each retained match to structured football data including match metadata, team statistics, lineups, formations, text event timelines, and pre-match prediction polls.

The dataset is designed for long-form sports video research where models need to connect full football broadcasts with structured temporal and match-level evidence.

---

Pipeline

The dataset is built through a multi-stage crawling and cleaning pipeline:

1. Playlist extraction: We started from the YouTube FIFA World Cup full-match playlist PLCGIzmTE4d0jq6wHT2TvSspZ_HLiIx4_y. From this playlist we extracted video IDs, URLs, titles, and team/year signals.
2. Match normalization: We parsed team names and World Cup editions from YouTube titles, normalizing aliases such as Korea Republic to South Korea and mapping editions such as 2022 Qatar, 2018 Russia, and 2002 Korea/Japan.
3. External match linking: Each video record was linked to a Soccer365 match page when a confident match page could be identified.
4. Structured data crawling: For linked matches, we crawled match metadata, stadium information, attendance, referees, lineups, formations, team statistics, text event timelines, and prediction poll results.
5. Folder-level packaging: Structured data was stored in one folder per match with a stable human-readable folder name.
6. Final validation: We checked file coverage across all structured match directories. Each structured match includes metadata.json, stats.json, events.json, lineups.json, and prediction.json.

---

Dataset Structure

FIFA_World_Cup_Games/
+-- meta.jsonl                              # One YouTube/source-link record per indexed video
+-- README.md
+-- images/
|   +-- poster.jpeg                        # Dataset poster
+-- games/
|   +-- 2022-final-Argentina-vs-France/
|   |   +-- metadata.json                  # Video link and match venue metadata
|   |   +-- stats.json                     # Team-level match statistics
|   |   +-- events.json                    # Text timeline / key events
|   |   +-- lineups.json                   # Starting XI, substitutes, formations
|   |   +-- prediction.json                # Pre-match prediction poll
|   +-- 2018-group-stage-Germany-vs-Mexico/
|   |   +-- metadata.json
|   |   +-- stats.json
|   |   +-- events.json
|   |   +-- lineups.json
|   |   +-- prediction.json
|   +-- ...
+-- poster_assets/
    +-- thumbnails/                        # YouTube thumbnails used to build the poster

Each match folder is named as:

YYYY-stage-TeamA-vs-TeamB

For example:

2022-final-Argentina-vs-France
2014-semi-finals-Brazil-vs-Germany
2018-group-stage-Germany-vs-Mexico

The repository does not include downloaded video files. Use the video_id or url field in meta.jsonl to retrieve videos independently when your use case and local policies allow it.

---

Dataset Overview

• Indexed YouTube video records: 40
• Structured match folders: 40
• World Cup editions covered: 7 editions, from 1998 France to 2022 Qatar
• Unique national teams: 25
• Structured files per retained match: metadata.json, stats.json, events.json, lineups.json, prediction.json
• Total event timeline rows: 3,543
• Event rows per match: 4 to 465, average 88.6
• Stages covered: group stage, 8th finals, quarter-finals, semi-finals, final
• Statistics fields observed: 22 team-level metric names
• Video distribution: Video files are not included; only YouTube IDs and URLs are provided

Edition Coverage

Edition	Indexed videos	Structured matches
2022 Qatar	8	8
2018 Russia	10	10
2014 Brazil	8	8
2010 South Africa	3	3
2006 Germany	5	5
2002 Korea/Japan	4	4
1998 France	2	2

Each indexed video has a corresponding structured match directory.

Stage Coverage

Stage	Structured matches
Group stage	20
8th finals	8
Quarter-finals	5
Semi-finals	2
Final	5

---

meta.jsonl

meta.jsonl is the top-level video/source index. Each line is a JSON object.

Field	Type	Description
video_id	string	YouTube video ID
url	string	YouTube watch URL
title	string	Original or lightly cleaned YouTube video title
teams	list[string]	Parsed teams in the match
world_cup	string	World Cup edition, such as 2022 Qatar
match_info_url	string	Linked Soccer365 match page

Example:

{
  "video_id": "HxBqMbI5kqQ",
  "url": "https://www.youtube.com/watch?v=HxBqMbI5kqQ",
  "title": "FULL MATCH: Brazil v Croatia | Quarter-Finals | FIFA WORLD CUP QATAR 2022",
  "teams": ["Brazil", "Croatia"],
  "world_cup": "2022 Qatar",
  "match_info_url": "https://soccer365.net/games/15292867/"
}

---

Match Metadata

Each metadata.json file links a source video to normalized match context.

Field	Type	Description
source_video.id	string	YouTube video ID
source_video.url	string	YouTube watch URL
source_video.title	string	Cleaned match title
source_video.date	string	Match date in YYYY-MM-DD format when available
match_info.home_team	string	Home/listed first team from the match source
match_info.away_team	string	Away/listed second team from the match source
match_info.stage	string	Tournament stage
match_info.datetime	string	Source datetime string
match_info.stadium	string	Stadium name
match_info.location	string	Stadium city and country
match_info.temperature	string	Source temperature string when available
match_info.weather	string	Weather text when available
match_info.viewers	string	Attendance
match_info.referees	list[string]	Referee crew when available
soccer365_url	string	Source match page

Example:

{
  "source_video": {
    "id": "ORzHdV_NVnQ",
    "url": "https://www.youtube.com/watch?v=ORzHdV_NVnQ",
    "title": "Argentina vs France -- 2022 FIFA World Cup Final",
    "date": "2022-12-18"
  },
  "match_info": {
    "home_team": "Argentina",
    "away_team": "France",
    "stage": "final",
    "datetime": "18.12.2022 23:59",
    "stadium": "Lusail",
    "location": "Lusail, Qatar",
    "temperature": "+35C",
    "weather": "",
    "viewers": "88,966",
    "referees": []
  },
  "soccer365_url": "https://soccer365.net/games/15292874/"
}

---

Match Statistics

Each stats.json file stores team-level match statistics. The top-level keys are statistic names, and each value maps team names to values.

Common statistic keys include:

Statistic	Description
Expected Goals (xG)	Expected goals when available, mainly for recent matches
Shots	Total shots
Shots on Target	Shots on target
Shots Blocked	Blocked shots
Saves	Goalkeeper saves
Possession %	Team possession percentage
Corners	Corner kicks
Fouls	Fouls committed
Offsides	Offside calls
Yellow Cards	Yellow cards
Red cards	Red cards
Attacks	Attacking sequences from the source page
Dangerous Attacks	Dangerous attacks from the source page
Passes	Total passes
Pass Accuracy %	Pass accuracy percentage
Free Kicks	Free kicks
Throw-ins	Throw-ins
Crosses	Crosses
Tackles	Tackles

Example:

{
  "Expected Goals (xG)": {
    "Argentina": "3.3",
    "France": "2.2"
  },
  "Shots": {
    "Argentina": "21",
    "France": "10"
  },
  "Shots on Target": {
    "Argentina": "9",
    "France": "5"
  }
}

The exact statistic set varies by match because older source pages expose fewer fields than recent matches.

---

Events

Each events.json file stores a source-derived text timeline. Rows are ordered chronologically when the source exposes a detailed report. Older games may only include a compact list of major match events.

Field	Type	Description
minute	string	Match minute or source marker
type	string	Source event type code, such as whistle, goal, subst, yc, rc, or empty string
description	string	Human-readable event text

Example:

[
  {
    "minute": "-",
    "type": "whistle",
    "description": "The referee starts the match"
  },
  {
    "minute": "1",
    "type": "",
    "description": "Mexico kick-off, and the game is underway."
  },
  {
    "minute": "7",
    "type": "",
    "description": "Good effort by Mats Hummels as he directs a shot on target, but the keeper saves it"
  }
]

This field should be treated as a text event timeline rather than a fully standardized optical tracking or official event feed.

---

Lineups

Each lineups.json file stores starting players, substitutes, and formations.

Team entries have the following structure:

Field	Type	Description
<team>.starting	list[object]	Starting XI players
<team>.substitutes	list[object]	Substitute bench players
number	string	Shirt number
name	string	Player name
formation.<team>	string	Formation, such as 4-3-3

Example:

{
  "Argentina": {
    "starting": [
      {"number": "23", "name": "Emiliano Martinez"},
      {"number": "10", "name": "Lionel Messi"}
    ],
    "substitutes": [
      {"number": "22", "name": "Lautaro Martinez"}
    ]
  },
  "France": {
    "starting": [
      {"number": "1", "name": "Hugo Lloris"},
      {"number": "10", "name": "Kylian Mbappe"}
    ],
    "substitutes": [
      {"number": "26", "name": "Marcus Thuram"}
    ]
  },
  "formation": {
    "Argentina": "4-3-3",
    "France": "4-2-3-1"
  }
}

---

Prediction Polls

Each prediction.json file stores the source page's pre-match user prediction poll.

Field	Type	Description
<team> win.percentage	string	Percentage of votes for a team win
<team> win.votes	string	Vote count for a team win
draw.percentage	string	Percentage of votes for a draw
draw.votes	string	Vote count for a draw

Example:

{
  "Argentina win": {
    "percentage": "36",
    "votes": "569"
  },
  "draw": {
    "percentage": "34",
    "votes": "525"
  },
  "France win": {
    "percentage": "30",
    "votes": "472"
  }
}

---

Quick Start

Install dependencies

pip install datasets huggingface_hub pandas yt-dlp

Load meta.jsonl locally

import json
from pathlib import Path

root = Path(".")

with open(root / "meta.jsonl", "r", encoding="utf-8") as f:
    videos = [json.loads(line) for line in f if line.strip()]

print(f"Loaded {len(videos)} video records")
print(videos[0])

Read one structured match

import json
from pathlib import Path

game_dir = Path("games/2022-final-Argentina-vs-France")

metadata = json.load(open(game_dir / "metadata.json", encoding="utf-8"))
stats = json.load(open(game_dir / "stats.json", encoding="utf-8"))
events = json.load(open(game_dir / "events.json", encoding="utf-8"))
lineups = json.load(open(game_dir / "lineups.json", encoding="utf-8"))
prediction = json.load(open(game_dir / "prediction.json", encoding="utf-8"))

print(metadata["source_video"]["url"])
print(stats["Shots"])
print(events[:3])
print(lineups["formation"])
print(prediction)

Load from Hugging Face

from datasets import load_dataset

# Replace repo_id with the final Hugging Face dataset repo name if different.
repo_id = "choucsan/FIFA_World_Cup_Games"

dataset = load_dataset(repo_id, data_files="meta.jsonl")
records = dataset["train"]

print(records[0])

Download a YouTube video independently

Videos are not redistributed in this dataset. To download a video for local research use, use the video_id or url field with yt-dlp:

yt-dlp -f "bv*+ba/b" \
  -o "games/2022-final-Argentina-vs-France/video/%(id)s.%(ext)s" \
  "https://www.youtube.com/watch?v=ORzHdV_NVnQ"

Please ensure that downloading and using videos complies with YouTube's Terms of Service, copyright rules, and your local research policies.

---

Applications

This dataset is intended for research on long-form sports video understanding and multimodal reasoning.

Long Video Understanding

• Full-match reasoning over football broadcasts
• Long-context video-language modeling
• Temporal localization of goals, substitutions, fouls, cards, and momentum shifts
• Understanding knockout matches, extra time, and penalty shootouts from long videos

Multimodal Question Answering

• Match-level QA combining video, team statistics, lineups, and event timelines
• Player and team questions grounded in starting XI and formation data
• Multi-hop questions linking event descriptions to final outcomes
• Venue, attendance, referee, and tournament-stage questions

Visual Retrieval

• Query-by-text retrieval of match moments
• Retrieval of goals, saves, cards, substitutions, corners, and attacks from event text
• Cross-modal retrieval between YouTube video frames, textual timelines, and structured statistics

Temporal Prediction

• Next-event prediction from match context and previous events
• Score and momentum progression modeling
• Pre-match expectation analysis using prediction polls and team metadata
• Comparing fan prediction distributions with actual match outcomes

Sports Analytics

• Team-level statistical comparison across World Cup editions
• Formation and lineup analysis
• Match intensity analysis using attacks, dangerous attacks, fouls, cards, and shots
• Historical comparison of classic World Cup matches

---

Download

Suggested Hugging Face dataset layout:

choucsan/FIFA_World_Cup_Games
+-- meta.jsonl
+-- games/
|   +-- */metadata.json
|   +-- */stats.json
|   +-- */events.json
|   +-- */lineups.json
|   +-- */prediction.json
+-- images/
    +-- poster.jpeg

Example commands:

# Download metadata only
huggingface-cli download choucsan/FIFA_World_Cup_Games \
  meta.jsonl \
  --repo-type dataset \
  --local-dir FIFA_World_Cup_Full_Match_Dataset

# Download all structured annotations
huggingface-cli download choucsan/FIFA_World_Cup_Games \
  --repo-type dataset \
  --local-dir FIFA_World_Cup_Full_Match_Dataset

Video files are not included. Use the url field in meta.jsonl or metadata.json to download videos independently when permitted.

---

Data Quality Notes

• The dataset links public YouTube full-match videos to Soccer365-derived structured match pages.
• events.json is a text timeline from the source page. It is not a frame-accurate official tracking feed.
• Event coverage varies by match. Recent matches may include dense commentary, while older matches may only expose key events.
• stats.json fields vary by match because the source exposes richer statistics for recent editions.
• meta.jsonl contains 40 indexed video records and games/ contains 40 structured match folders.
• Each structured match folder has been checked against meta.jsonl by YouTube video ID, Soccer365 URL, and participating teams.
• Team order follows the source pages and should not always be interpreted as FIFA's official home/away designation.

---

Contact

For questions, corrections, or collaboration requests:

choucisan@gmail.com