ArTyDi-QA

ArTyDi-QA is a dataset for Modern Standard Arabic (MSA) designed for Question Answering (QA) and Question Generation (QG) tasks. It is derived from the Arabic subset of the TyDi QA benchmark and adapted into a SQuAD 2.0-style extractive format.

The dataset features natively authored, information-seeking questions and incorporates paragraph-level unanswerable questions.

Structure

The dataset is organized into two compressed archives. All data is provided in standard JSONL (JSON Lines) format.

QA.zip

Contains the Extractive QA dataset. This subset includes both answerable and unanswerable questions.

• train.jsonl
• val.jsonl
• test.jsonl

QG.zip

Contains the Question Generation dataset. This subset consists strictly of answerable context-question pairs, ideal for training generative models.

• train.jsonl
• val.jsonl
• test.jsonl

Data Format

Each line in the .jsonl files represents a single data example as a valid JSON object with the following fields:

Field	Type	Description
id	String	A unique identifier for the example.
title	String	The title of the Wikipedia article from which the context was extracted.
context	String	The paragraph text providing the information context.
question	String	The question text relating to the context.
answers	Object	An object containing the ground truth answer(s). See structure below.

The answers Object

The answers field is a dictionary containing lists, allowing for multiple valid answers or spans:

• text (List[String]): A list of acceptable answer strings.
• start_char (List[Int]): A list of the starting character indices for the answer spans in the context.
• end_char (List[Int]): A list of the ending character indices for the answer spans in the context.

Special Answer Types

Unanswerable Questions In the QA subset, approximately 29% of the questions are unanswerable based on the provided paragraph. For these examples:

• text: Contains the string "لا يمكن الإجابة على السؤال من النص" (Cannot answer the question from the text).
• start_char and end_char: Set to [-1].

Yes/No Questions For binary questions where the answer is implicit in the text but not a specific span:

• text: Contains either "نعم" (Yes) or "لا" (No).
• start_char and end_char: Set to [-1].

Statistics

Statistic	ArTyDi-QA	ArTyDi-QG
Total Examples	24,120	17,183
Train Examples	18,272	12,962
Validation Examples	4,523	3,223
Test Examples	1,325	998
Yes/No Questions	6%	8.5%
Unanswerable	29%	N/A