README.md

metadata

license: odc-by
task_categories:
  - question-answering
language:
  - en
pretty_name: ASTRA-QA
size_categories:
  - 1K<n<10K
configs:
  - config_name: questions
    default: true
    data_files:
      - split: train
        path: questions.jsonl
  - config_name: corpus
    data_files:
      - split: train
        path: corpus.jsonl

ASTRA-QA: A Benchmark for Abstract Question Answering over Documents

ASTRA-QA, short for AbSTRAct Question Answering over documents, is a dataset and benchmark for document-level, synthesis-heavy question answering in retrieval-augmented generation systems. It evaluates whether a system can read long documents, organize evidence, and produce grounded abstractive answers with reference-based assessment, rather than only retrieve short facts.

Dataset Summary

• 869 questions in total
• 5 task types: Single-Sum, Pair-Comp, Multi-Comp, Enum, and Temp
• 2 source domains: academic documents and news documents
• 3 retrieval scopes used in the benchmark: Simple, Middle, and Hard

Dataset Files

• corpus.jsonl: source documents used for retrieval and evidence grounding
• questions.jsonl: abstractive question-answer pairs with topic-set style answers and benchmark metadata

Task Types

• Single-Sum: summarize a single document into a compact grounded answer
• Pair-Comp: compare two documents, methods, entities, or events
• Multi-Comp: synthesize comparisons across multiple targets
• Enum: enumerate key items, themes, findings, or contributions
• Temp: reconstruct temporally evolving events over a time window

Data Sources

ASTRA-QA is constructed from publicly available sources, including arXiv, OpenReview, and news articles collected through mediastack.com.

License

This dataset is released under the Open Data Commons Attribution License (ODC-By).

The ODC-By license applies to the dataset annotations, organization, metadata, and benchmark construction. Original source documents remain subject to their respective licenses and terms of use. Users are responsible for complying with the original licenses and source-specific usage terms when using the source content.

Citation

Citation information will be added when the paper is released.