README.md

metadata

configs:
  - config_name: engineering_human_factor
    data_files:
      - split: train
        path: data/engineering_human_factor.json
  - config_name: material_science_additive_manufacturing
    data_files:
      - split: train
        path: data/material_science_additive_manufacturing.json
  - config_name: physics_surface_enhanced_raman_spectroscopy
    data_files:
      - split: train
        path: data/physics_surface_enhanced_raman_spectroscopy.json
  - config_name: public_health_infectious_disease_modeling
    data_files:
      - split: train
        path: data/public_health_infectious_disease_modeling.json
  - config_name: earth_science_remote_sensing
    data_files:
      - split: train
        path: data/earth_science_remote_sensing.json
pretty_name: Intrabench
task_categories:
  - question-answering
  - text-classification
size_categories:
  - n<1K
license: mit
language:
  - en

IntraBench Spaceholding

This dataset is a five-domain benchmark of multiple-choice research questions derived from scientific literature. Each example represents one question-paper pair with a gold letter answer and supporting metadata.

Subsets

• engineering_human_factor: Engineering - Human Factor
• material_science_additive_manufacturing: Material Science - Additive Manufacturing
• physics_surface_enhanced_raman_spectroscopy: Physics - Surface Enhanced Raman Spectroscopy
• public_health_infectious_disease_modeling: Public Health - Infectious-disease Modeling
• earth_science_remote_sensing: Earth Science - Remote Sensing

Data Schema

Each record contains:

• subject: human-readable subject name
• paper_id: paper index within the subject subset
• paper_title: canonical benchmark paper title
• question_id: question identifier
• question: canonical benchmark question wording
• choices: answer choice dictionary keyed by A-F
• answer: gold answer letter
• metadata: extra metadata with:
  • Task-oriented Category
  • question_key_term
  • term_explanation

Notes

• The published dataset excludes source PDFs, markdown paper extracts, and internal evaluation outputs.
• The benchmark uses canonical paper titles and normalized question wording across all five domains.
• Data files are stored as JSON arrays and exposed as separate Hugging Face configs.

Citation

If you publish this dataset, add the benchmark paper citation here.