Social-AI-2026/worldcup2026

⚽ WorldCup Arena

A Leakage-Free Forecasting Benchmark on a Live Tournament

Can a language model forecast a match — when the match had not been played at the moment it was asked?

---

  🌐   Language / 语言  :  中文   ▾  

📊 四张表

点开本页顶部的 Data Studio 标签即可浏览，也可以直接按名字加载。

Config	行数	内容
fixtures	104	基准本体 —— 喂给模型的头部信息，以及结算后的 90 分钟赛果，七个盘口全部推导好（outcome_1x2、over_2_5、both_score、odd_total）
dossiers	2,208	简报索引 —— 46 快照 × 48 队，每份的篇幅。正文本身在 dossiers/ 目录下，不在表里；某份简报的位置是 dossiers/<快照>/<队伍代码>/summary.md。
teams	48	球队元数据，中英双名
venues	16	场馆，含海拔（米）

from datasets import load_dataset

fx = load_dataset("Social-AI-2026/worldcup2026", "fixtures", split="test")

表格是概览，不是语料本身。权威文件仍是 core/ 与 dossiers/ 下的 JSON 和 Markdown —— 算分代码读的是后者，snapshot_download 拉的也是它们。

---

⚡ 数据集简介

这里是 2026 世界杯的简报材料与标准答案，作为一个预测基准发布。

这里的每一场比赛，都是在它的简报材料冻结之后才踢的。模型不可能从预训练里背到答案，因为提问的那一刻，答案在世上任何地方都不存在。这里没有匿名化步骤、没有时间截断、没有泄露检测器——因为一样都不需要。数据污染在这里不是被"缓解"了，而是构造上就不可能发生。

本数据集不含我们自己的任何预测。它是卷子和标准答案，不是谁的成绩单。

• 主页 / 代码 —— https://github.com/co-minder/worldcup2026-codebase
• 字段级规范 —— SCHEMA.md
• 语言 —— 档案正文为中文（zh）；赛程、球队、场馆元数据为中英双语
• 联系方式 —— 见 CITATION.cff

时间跨度。 2026 年 6 月 10 日 – 7 月 19 日，全部 104 场，小组赛到决赛。

🎯 任务

模型被放在某场比赛开球之前。它能看到那一刻为止的档案，看不到之后的任何东西，然后为即将开踢的这场填一张预测卡：

盘口	问什么	选项	分值
让球	热门能不能打赢固定盘口线？	主队赢盘 / 走盘 / 客队赢盘	4
半全场	半场谁领先，全场谁赢？	9 种组合	3
胜平负	谁赢？	主胜 / 平 / 客胜	2
大小 2.5	总进球是否超过两个？	大 / 小	2
双方进球	两队是否都进球？	是 / 否	2
正确比分	具体比分是多少？	X–Y，开放作答	2
单双	总进球是单数还是双数？	单 / 双	1

七项一律按 90 分钟比分结算（含补时）。加时和点球一概不算——淘汰赛 90 分钟战平，在这里就记平局，不管谁晋级。决赛 ESP_vs_ARG 记作 0-0 正是这个原因。

另有两类题在开幕前只问一次：12 个小组各自的头名，以及一个全局彩池——冠军、决赛双方、四强、夺冠大洲，以及整届总进球是否超过 285.5 这条线。

平凡基线**不是 50%**：机械跟着博彩盘口热门押胜平负，准确率是 **64.4%**。配套报告里六个前沿模型的均值是 **63.9%**——那条线才是要打败的对象，而它们一个都没打败。

🗂️ 结构

core/
    matches.json    104 场赛程 —— 轮次、小组、日期、开球时间、场馆、双方
    teams.json      48 支球队 —— 代码、队名、所属大洲
    venues.json     16 座球场 —— 城市、海拔、草皮、顶棚、坐标
    results.json    标准答案 —— 104 场比分、半场比分、
                    12 个小组头名、全局彩池

dossiers/
    <YYYY-MM-DD_HHMM>/<队伍代码>/summary.md      模型可以看到的简报
    <YYYY-MM-DD_HHMM>/<队伍代码>/CHANGELOG.md    这一版档案改了什么、什么时候改的

一个快照是那一分钟世界的完整状态，不是增量。<队伍代码> 是三字母 FIFA 代码，也是贯穿全部文件的连接键。每个文件的每个字段都记在 SCHEMA.md 里。

让这个基准站得住的唯一一条规矩：取时间戳严格早于该场 kickoff_ts 的最近一版快照。再晚的快照可能已经含有你正要问的那场的结果。examples/load_and_grade.py 把这条规矩写成了代码。

📊 规模

轮次	场数	平局	涉及球队
小组赛	72	20	48
32 强	16	3	32
16 强	8	1	16
八强	4	2	8
半决赛	2	0	4
决赛与季军赛	2	1	4
合计	104	27	48

	规模
比赛	104 场 —— 整届一场不落
每场盘口	7 个，权重 4/3/2/2/2/2/1
整届层面的题	12 个小组头名 + 6 项全局彩池
球队档案	48 队 × 46 个时间戳快照
档案体量	赛事后期每队约 20–30 kB
时间跨度	连续 39 天，2026-06-10 至 07-19
总进球	303 球，赛前盘口线 285.5

按 90 分钟比分统计，整届是 47 场主胜、27 场平局、30 场客胜。

🚀 快速开始

跑示例

pip install huggingface_hub
hf download Social-AI-2026/worldcup2026 --repo-type dataset --local-dir ./worldcup2026
cd worldcup2026
python3 examples/load_and_grade.py --match M64

它会挑出这场比赛、解析出唯一合法的那一版快照、按模型实际看到的样子拼好简报，并打印标准答案与逐个盘口已推导好的结果。

直接读文件

import json, pathlib
from huggingface_hub import snapshot_download

root = pathlib.Path(snapshot_download(
    repo_id="Social-AI-2026/worldcup2026",
    repo_type="dataset",
    allow_patterns=["core/*", "dossiers/2026-06-10_1310/**"],
))

matches = json.loads((root / "core/matches.json").read_text(encoding="utf-8"))
results = json.loads((root / "core/results.json").read_text(encoding="utf-8"))

fixture = matches[0]                                     # M1 · MEX vs RSA · 2026-06-11
brief = "\n\n".join(
    (root / f"dossiers/2026-06-10_1310/{t}/summary.md").read_text(encoding="utf-8")
    for t in (fixture["team_a"], fixture["team_b"])      # 该快照早于开球
)

# ...把 brief 喂给你的模型，让它回答七个盘口...

print(results["matches"]["MEX_vs_RSA"])                  # {'score': '2-0', 'ht': '1-0'}

只想要 508 kB 的元数据、不下那 164 MB 档案，就传 allow_patterns=["core/*"]。

不要把 core/results.json 喂给被测模型。 那是标准答案，公开出来是为了让结果可被审计。

复现报告

git clone https://github.com/co-minder/worldcup2026-codebase
cd worldcup2026-codebase
hf download Social-AI-2026/worldcup2026 --repo-type dataset --local-dir ./wc_runs
python3 main.py --score

🔬 制作过程

为什么这么做。 预测类基准有一个测量上的死结：在真实的、已经揭晓的事件上，强模型可能是在回忆而不是在预测；而现有的所有补救手段——匿名化、不断换新题、泄露检测——都只是代理，真正想要的那个性质是"提问时答案根本不存在"。一届还没踢的赛事恰好精确地给出这个性质，而世界杯给得很密：同一领域、连续 39 天、104 场客观可判定的比赛。

数据来源。 档案每日汇编，每一条写入前都用至少一个独立来源交叉核对：Wikipedia（阵容、世界杯历史、预选赛）、Transfermarkt（身价、年龄、俱乐部）、eloratings.net（Elo）、FIFA（官方赛程、排名、赛果）、以及公开赛事报道（逐场叙事与逐分钟时间线）。

档案构建是一条流水线，不是手工：

1. 采集 —— 一次采集产出一个完整的时间戳快照；原始页面在任何加工之前先逐字归档。
2. 蒸馏 —— 一条智能体链把原始条目压进档案的各个章节，硬数据（阵容、身价、战绩）与叙事分开存放。
3. 过闸 —— 每条候选内容进档案前都要过三道判断：是新的吗？重要吗？是对的吗？被拒的条目按队记录，并写明拒因。

标注。 常规意义上的标注不存在。标准答案是 FIFA 的官方 90 分钟比分，不是人工打标。淘汰赛若在加时或点球决出，记录的仍是 90 分钟比分，并在该条目上加一条 note。

隐私。 档案里出现的球员、主帅、主裁均为职业身份下的公众人物，记录的每一项事实（阵容、身价、出场、纪律记录）都取自对其工作的公开报道。不含任何私人个体、联系方式或非公开信息。

⚠️ 使用须知

预期用途。 在"记忆不可能奏效"的条件下评测大模型与智能体系统的预测能力。报告成绩时请一并写明模型的训练截止时间。

没有包含什么，以及为什么：

• 我们自己的预测、积分榜、赛后复盘 —— 本发布是卷子和标准答案。把某一家实验室的答案和它们摆在一起，等于邀请别人朝答案调参。
• 模型原始回答与存档 prompt —— 同上；这些已保留，可应要求提供以供审计。
• 采集过程抓到的原始 HTML —— 属于采集过程而非实验记录，报告里没有任何数字依赖它。
• 新闻总结线与播报线的中间文件 —— 流水线的中间产物，不是基准材料。

局限。

• 只有一届赛事。104 场既是这届的全部，也是"所有世界杯"这个总体里的一个样本。要缩小不确定性只能再跑一届，不是把这一届算得更巧。
• 无污染的保证有时效。它只对训练数据早于 2026 年 7 月的模型成立。更晚的模型可能已经读过赛果，而本数据集无法检测这一点。
• 答案是分类选项，不是概率。七个盘口都是单选，裸跑出来算不了 proper scoring rule 和校准曲线。要校准就自己让模型输出概率：材料支持这么做，标准答案不要求。
• 档案正文是中文。元数据是双语的，简报正文不是；中文弱的模型会因为与预测能力无关的原因吃亏。
• 快照按分钟而非按场次。一天有多场时，一版快照服务所有场次；它绝不含被问的那一场，但可能含同一天更早的比赛——这是合法的，那场已经踢完了。
• 覆盖度不均，和足球报道本身一样不均。欧洲与南美球队的公开记录更厚，档案也更丰富。身价是某一家商业机构的估值。每场记录的让球盘是庄家的判断，看到它的模型可能被锚定——这个效应报告里量化了，没有假装不存在。

不构成投注建议。 博彩赔率出现在这里，是因为它是公开可得的、对一场足球比赛最锐利的预测，因而是诚实的对照基线。本数据集中没有任何内容是在建议赌博。

发现问题？ 比分记错、快照日期不对、或者某份档案在那个时间戳上不该有的信息出现了——这都是值得报的 bug。在本仓库开一个 discussion，或去代码仓开 issue，带上场次键和文件路径。

📄 引用

@techreport{worldcuparena2026,
  title  = {{WorldCup Arena}: Prospective, Leakage-Free Evaluation of
            Frontier LLMs on a Live Tournament},
  author = {Wang, Zhenran and Bian, Zhonghan and Li, Jinsong and Qi, Zhangyang},
  year   = {2026},
  note   = {\url{https://github.com/co-minder/worldcup2026-codebase}}
}

机器可读的元数据在 CITATION.cff。版本变更记在 CHANGELOG.md。

⚖️ 许可

数据采用 CC BY 4.0。底层事实与任何被引用素材的权利归各自权利人所有。构建与评分它的流水线以 MIT 许可单独发布在 co-minder/worldcup2026-codebase。

  🌐   Language / 语言  :  English   ▾  

📊 The four tables

Open the Data Studio tab at the top of this page to browse them, or load one by name.

Config	Rows	What it is
fixtures	104	The benchmark — the header a model was given, and the settled 90-minute result with every market derived (outcome_1x2, over_2_5, both_score, odd_total)
dossiers	2,208	An index of the briefings — 46 snapshots × 48 teams, with each one's length. The prose itself lives under dossiers/, not in the table; a briefing is at dossiers/<snapshot>/<team>/summary.md.
teams	48	Team metadata, bilingual names
venues	16	Stadiums, including altitude in metres

from datasets import load_dataset

fx = load_dataset("Social-AI-2026/worldcup2026", "fixtures", split="test")

The tables are an overview, not the corpus. The authoritative files stay the JSON and Markdown under core/ and dossiers/ — those are what the scoring code reads, and what snapshot_download fetches.

---

⚡ Dataset description

These are the briefing materials and the answer key for the 2026 FIFA World Cup, released as a forecasting benchmark.

Every match here was played after its briefing material was frozen. A model cannot have memorised the answer from pre-training, because at the moment of asking the answer did not exist anywhere. There is no anonymisation step, no cutoff to enforce and no leak detector, because none of them are needed. Contamination is not mitigated — it is impossible by construction.

This release contains no predictions of our own. It is the question paper and the answer key, not anyone's score.

• Homepage / code — https://github.com/co-minder/worldcup2026-codebase
• Field-level schema — SCHEMA.md
• Languages — dossier prose is Chinese (zh); fixture, team and venue metadata is bilingual
• Point of contact — see CITATION.cff

Period. 10 June – 19 July 2026, all 104 matches, group stage through the final.

🎯 The task

A model is placed before a kickoff. It sees the dossiers as they stood at that moment and nothing later, then fills in a prediction card for the match that is about to be played:

Market	Question	Options	Points
Handicap	Does the favourite cover the fixed line?	home covers / push / away covers	4
Half-time / full-time	Who leads at the break, who wins?	9 combinations	3
Match outcome (1X2)	Who wins?	home / draw / away	2
Over/under 2.5	More than two goals?	over / under	2
Both teams to score	Do both sides score?	yes / no	2
Correct score	What is the scoreline?	X–Y, open	2
Odd/even	Is the goal total odd or even?	odd / even	1

All seven settle on the 90-minute scoreline, including stoppage time. Extra time and penalties never count — a knockout tie level after 90 minutes is a draw here, whoever advanced. The final, ESP_vs_ARG, is recorded as 0-0 for exactly this reason.

Two further question types are asked once, before the opening match: the winner of each of the 12 groups, and an outright pool — champion, both finalists, the four semi-finalists, the winning confederation, and whether tournament goals exceed a line of 285.5.

The trivial anchor is not 50%: mechanically backing the side the bookmaker's line favours scores 64.4% on match outcome. In the accompanying report, six frontier LLMs average 63.9% — that baseline is the number to beat, and none of them did.

🗂️ Structure

core/
    matches.json    104 fixtures — round, group, date, kickoff, venue, both teams
    teams.json      48 teams — codes, names, confederation
    venues.json     16 stadiums — city, altitude, surface, roof, coordinates
    results.json    the answer key — 104 scorelines, half-time scores,
                    12 group winners, the outright pool

dossiers/
    <YYYY-MM-DD_HHMM>/<TEAM>/summary.md      the briefing a model may see
    <YYYY-MM-DD_HHMM>/<TEAM>/CHANGELOG.md    what changed in that dossier, and when

A snapshot is a complete state of the world at that minute, not a diff. <TEAM> is the three-letter FIFA code, the key that ties everything together. Every field of every file is documented in SCHEMA.md.

The one rule that keeps this benchmark honest: use the latest snapshot whose timestamp is strictly earlier than the fixture's kickoff_ts. A later snapshot may already contain the result you are about to ask about. examples/load_and_grade.py implements the rule.

📊 Scale

Round	Matches	Draws	Teams involved
Group stage	72	20	48
Round of 32	16	3	32
Round of 16	8	1	16
Quarter-finals	4	2	8
Semi-finals	2	0	4
Final & third place	2	1	4
Total	104	27	48

	Scale
Matches	104 — every fixture of the tournament, none skipped
Markets per match	7, weighted 4/3/2/2/2/2/1
Tournament-level questions	12 group winners + a 6-slot outright pool
Team dossiers	48 teams × 46 timestamped snapshots
Dossier size	~20–30 kB per team late in the tournament
Period	39 consecutive days, 10 June – 19 July 2026
Goals scored	303, against a pre-tournament line of 285.5

Outcomes are 47 home wins, 27 draws and 30 away wins on the 90-minute scoreline.

🚀 Quick start

Run the example

pip install huggingface_hub
hf download Social-AI-2026/worldcup2026 --repo-type dataset --local-dir ./worldcup2026
cd worldcup2026
python3 examples/load_and_grade.py --match M64

It picks the fixture, resolves the only legal snapshot for it, assembles the briefing exactly as a model would see it, and prints the answer key with each market already derived.

Load the files directly

import json, pathlib
from huggingface_hub import snapshot_download

root = pathlib.Path(snapshot_download(
    repo_id="Social-AI-2026/worldcup2026",
    repo_type="dataset",
    allow_patterns=["core/*", "dossiers/2026-06-10_1310/**"],
))

matches = json.loads((root / "core/matches.json").read_text(encoding="utf-8"))
results = json.loads((root / "core/results.json").read_text(encoding="utf-8"))

fixture = matches[0]                                     # M1 · MEX vs RSA · 2026-06-11
brief = "\n\n".join(
    (root / f"dossiers/2026-06-10_1310/{t}/summary.md").read_text(encoding="utf-8")
    for t in (fixture["team_a"], fixture["team_b"])      # snapshot precedes kickoff
)

# ... ask your model for the seven markets, given `brief` ...

print(results["matches"]["MEX_vs_RSA"])                  # {'score': '2-0', 'ht': '1-0'}

Pass allow_patterns=["core/*"] to fetch only the 508 kB of metadata without the 164 MB of dossiers.

Do not feed core/results.json to the model under evaluation. It is the answer key, released so that results can be audited.

Reproduce the report

git clone https://github.com/co-minder/worldcup2026-codebase
cd worldcup2026-codebase
hf download Social-AI-2026/worldcup2026 --repo-type dataset --local-dir ./wc_runs
python3 main.py --score

🔬 Creation

Curation rationale. Benchmarks for forecasting run into a measurement problem: on real, resolved events a strong model may be recalling rather than predicting, and every existing remedy — anonymisation, refreshed question pools, leak detection — is a proxy for the property one actually wants, that the answer did not exist when the question was asked. A tournament that has not been played gives that property exactly, and a World Cup gives it densely: 104 objectively adjudicated fixtures in one domain over 39 consecutive days.

Source data. Dossiers were assembled daily and cross-checked against at least one independent source before any line was written in: Wikipedia (squads, tournament history, qualification), Transfermarkt (market values, ages, clubs), eloratings.net (Elo), FIFA (official fixtures, rankings and results), and public match reporting (the per-match narrative and minute-by-minute timeline).

Dossier construction runs as a pipeline, not by hand:

1. Collect — one collection pass produces one complete timestamped snapshot; raw pages are archived verbatim before any processing.
2. Distil — an agent chain compresses raw items into the dossier's sections, keeping hard data (squad, values, results) separate from narrative.
3. Gate — every candidate line passes a three-way check before it enters a dossier: is it new? does it matter? is it correct? Rejections are logged per team with the reason.

Annotations. There are none in the usual sense. Ground truth is the official 90-minute scoreline, taken from FIFA, not from human labelling. Where a knockout tie was decided in extra time or on penalties, the recorded result is still the 90-minute one and the entry carries a note.

Personal and sensitive information. The dossiers name footballers, coaches and referees — all public figures acting in a professional capacity, and every recorded fact (squad membership, market value, appearances, disciplinary record) comes from public reporting about their work. No private individuals, contact details or non-public information are included.

⚠️ Considerations

Intended use. Evaluating the forecasting ability of LLMs and agent systems in a setting where memorisation is impossible. Report the model's training cutoff alongside any score.

What is not included, and why:

• Our own predictions, scorecard and post-match reviews — this release is the question paper and the answer key. Publishing one lab's answers alongside them invites tuning against them.
• Raw model responses and archived prompts — same reason; they are retained and available for audit on request.
• The raw HTML captured during collection — collection process rather than experimental record; nothing in the report depends on it.
• The news-summary and broadcast working files — intermediate artefacts of the pipeline, not benchmark material.

Limitations.

• One tournament. There is exactly one 2026 World Cup. The 104 matches are the entire population of this event and simultaneously a sample of size one from the space of tournaments. Uncertainty narrows by running another tournament, not by re-analysing this one more cleverly.
• The contamination guarantee expires. It holds for models whose training data predates July
  2026. A later model may have read the results, and this dataset cannot detect that.
• Categorical answers, not probabilities. The seven markets take a single pick, so no proper scoring rule and no calibration curve follow from a plain run. Elicit probabilities yourself if you want them — the material supports it, the answer key does not require it.
• Dossier prose is Chinese. Metadata is bilingual, the briefing is not; a model weak in Chinese is handicapped for reasons unrelated to forecasting.
• Snapshots are per-minute, not per-fixture. On a day with several matches one snapshot serves them all. It never contains the fixture being asked about, but it may contain an earlier match from the same day.
• Coverage is uneven in the way football reporting is uneven. European and South American squads have deeper public records, so their dossiers are richer. Market values are one commercial vendor's estimates. The handicap line is a bookmaker's opinion, and a model shown it may anchor on it — an effect the report measures rather than assumes away.

Not betting advice. Bookmaker odds appear here because they are the sharpest publicly available forecast of a football match, and therefore the honest baseline to beat. Nothing in this dataset is a recommendation to gamble.

Found a problem? A wrong scoreline, a mis-dated snapshot, or a dossier containing information it should not have had at that timestamp is a bug worth reporting. Open a discussion here, or an issue in the code repository, with the fixture key and the file.

📄 Citation

@techreport{worldcuparena2026,
  title  = {{WorldCup Arena}: Prospective, Leakage-Free Evaluation of
            Frontier LLMs on a Live Tournament},
  author = {Wang, Zhenran and Bian, Zhonghan and Li, Jinsong and Qi, Zhangyang},
  year   = {2026},
  note   = {\url{https://github.com/co-minder/worldcup2026-codebase}}
}

Machine-readable metadata is in CITATION.cff. Revisions are listed in CHANGELOG.md.

⚖️ Licence

Data under CC BY 4.0. Rights in the underlying facts and in any quoted source material remain with their respective holders. The pipeline that builds and scores it is released separately under the MIT Licence at co-minder/worldcup2026-codebase.