sooo66/semeval2026-task12-dataset

SemEval-2026 Task 12: Abductive Event Reasoning

This repository hosts the public release of the dataset for SemEval-2026 Task 12, Abductive Event Reasoning (AER). The task evaluates whether a system can identify the most plausible direct cause of a target event from a set of multiple-choice candidates, using retrieved news context.

The dataset is published with two configurations:

• questions: multiple-choice reasoning instances.
• docs: contextual topic-level document collections used to support reasoning.

Dataset Summary

• Train questions: 1,819
• Validation questions: 400
• Test questions: 612
• Sample questions: 200
• Train topics: 36
• Validation topics: 36
• Test topics: 24
• Sample topics: 10

In this public release, the test split includes gold labels under golden_answer. These labels were merged from the internal reference file used during task development, so this release should be treated as a fully labeled public benchmark rather than a blind evaluation set.

Repository Structure

.
├── docs
│   ├── sample.json
│   ├── test.json
│   ├── train.json
│   └── validation.json
└── questions
    ├── sample.jsonl
    ├── test.jsonl
    ├── train.jsonl
    └── validation.jsonl

questions files are line-delimited JSON (.jsonl). docs files are JSON arrays, one topic record per element.

Question Format

Each line in questions/*.jsonl is a JSON object with the following fields:

{
  "topic_id": 11,
  "id": "q-101",
  "target_event": "Short description of an observed event.",
  "option_A": "Candidate explanation A.",
  "option_B": "Candidate explanation B.",
  "option_C": "Candidate explanation C.",
  "option_D": "Candidate explanation D.",
  "golden_answer": "A,B"
}

Field notes:

• id: unique instance identifier.
• topic_id: identifier linking the question to its contextual documents.
• target_event: observed event to explain.
• option_A to option_D: candidate direct causes.
• golden_answer: one or more correct labels, comma-separated.

Context Format

Each record in docs/*.json contains a topic and its retrieved supporting documents:

{
  "topic_id": 11,
  "topic": "OpenAI releases ChatGPT.",
  "docs": [
    {
      "title": "Article title",
      "id": "doc-001",
      "link": "https://example.com",
      "snippet": "Short summary of the document.",
      "source": "News source",
      "imageUrl": "Base64imageUrl",
      "content": "Full document text."
    }
  ]
}

Usage

from datasets import load_dataset

questions = load_dataset("YOUR_NAMESPACE/semeval2026-task12-dataset", "questions")
docs = load_dataset("YOUR_NAMESPACE/semeval2026-task12-dataset", "docs")

Notes

• The train and validation splits use different question sets over the same 36 topic-level document collections.
• The public test split is labeled.
• The original task evaluation was hosted on Codabench: https://www.codabench.org/competitions/12446/

Licensing And Use

This dataset card uses license: other as a placeholder. Update the license metadata if a more specific redistribution license applies to the task release and the included document content.