Dataset Card for long_squad_v2

long_squad_v2 is a long-context question answering dataset based on the SQuAD v2 format. It was constructed by concatenating multiple SQuAD v2 contexts to significantly increase the average document length, enabling training and evaluation of models on long-range understanding and sparse answer retrieval tasks.

Dataset Details

Uses

To load the dataset using the 🤗 Datasets library:

from datasets import load_dataset

dataset = load_dataset("huutuan/long_squad_v2")

Direct Use

This dataset is ideal for:

• Training and benchmarking long-context QA models such as Longformer, BigBird, and Claude
• Evaluating retrieval-augmented generation systems
• Research in sparse answer extraction and long-range attention

Out-of-Scope Use

• Not suitable for short-context QA tasks without adaptation
• Not a real-world long document corpus (contexts are concatenated artificially)

Dataset Structure

Each sample is structured as:

• id: Unique identifier
• title: Document title
• context: A long paragraph composed of multiple original SQuAD v2 contexts
• question: A natural language question
• answers: Dictionary with answer_start (char index) and text (list of answers)

Split statistics:

	Train	Validation
Total samples	130,319	11,873
Samples with answers	86,821	5,928
Samples without answers	43,498	5,945
Avg. context length (words)	5,918.43	6,330.12
Median context length	6,112	6,178
Std. context length	995.41	1,407.92
Min context length	2,908	4,190
Max context length	8,390	12,024
Avg. question length	9.89	10.02
Median question length	9	10
Std. question length	3.42	3.49
Avg. answer length	3.16	3.04
Median answer length	2	2
Std. answer length	3.39	2.62

Dataset Creation

Curation Rationale

To enable evaluation of models on extremely long contexts, which better reflect real-world documents (e.g., research papers, manuals, legal documents).

Source Data

Data Collection and Processing

• Based on original SQuAD v2 dataset
• Contexts are concatenated across articles while preserving original question-answer mappings
• Ensured no overlap between train and validation

Who are the source data producers?

Original data comes from Wikipedia articles curated by the Stanford NLP group (SQuAD v2).

Bias, Risks, and Limitations

• Dataset is synthetic in structure due to context concatenation
• Models trained on this data may not generalize to real-world long documents without further tuning
• Some questions may become harder to answer due to diluted relevance in longer contexts

Recommendations

Users should evaluate models carefully, especially on real-world long-context QA settings. Consider supplementing this dataset with actual long document datasets.