CogBench: A Benchmark for Evaluating Cognitive-Level Control in LLM Question Generation

Dataset Description

CogBench is a benchmark dataset for evaluating whether large language models can generate educational questions at specific cognitive levels according to Bloom's Taxonomy (Anderson & Krathwohl, 2001).

Dataset Summary

• 27,099 questions across 16 academic subjects
• 6 cognitive levels (Remember, Understand, Apply, Analyze, Evaluate, Create)
• 739 human-labeled questions from peer-reviewed sources
• 26,360 silver-labeled questions using the CCS classifier (82% accuracy)
• 4 data sources: Yahya (2012), Mohammed & Omar (2020), OpenStax QA, OpenStax MCQ

Supported Tasks

• Bloom's Taxonomy Classification: Predict the cognitive level (1-6) of a given question
• Cognitive-Level-Controlled Question Generation: Evaluate whether an LLM can generate questions at a specified Bloom's level

Languages

English

Dataset Structure

Data Fields

Field	Type	Description
question_id	string	Unique identifier
question_text	string	The question text
question_type	string	open_ended or mcq
subject	string	Academic subject (e.g., biology, physics)
source	string	Data source identifier
bloom_level	int	Bloom's Taxonomy level (1-6)
bloom_name	string	Level name (Remember, Understand, Apply, Analyze, Evaluate, Create)
label_source	string	How the label was assigned (original_author, taxonomy_mapping, or ccs_phase1_silver)
confidence	float	CCS model confidence score (for silver labels)

Data Splits

Split	N	Description
train	~21,800	Silver-labeled + human-labeled (for training classifiers)
validation	~2,600	Silver-labeled (for tuning)
test	~2,600	Silver-labeled (for evaluation)
human_annotated	739	Human-labeled from peer-reviewed sources (gold standard)

Subject Distribution

Subject	Count	Subject	Count
Physics	5,294	Economics	1,257
Mathematics	4,665	Computer Science	1,001
Biology	3,421	Political Science	824
Business	2,653	Astronomy	781
Chemistry	2,326	History	749
Psychology	1,665	Sociology	736
Nursing	610	Philosophy	228
Anthropology	150	General	739

Dataset Creation

Source Data

1. Yahya (2012): 600 questions manually classified by education researchers. Published in PLOS ONE.
2. Mohammed & Omar (2020): 141 questions with expert Bloom's classifications. Published in PLOS ONE.
3. OpenStax QA: ~9,900 question-answer pairs from open-source textbooks via HuggingFace.
4. OpenStax MCQ: ~16,400 multiple-choice questions scraped from OpenStax interactive content.

Annotations

• Human labels (739 questions): Original Bloom's classifications from peer-reviewed publications
• Silver labels (26,360 questions): Assigned by CCS (Cognitive Classification Score), a fine-tuned BERT-base-uncased model achieving 82.0% exact accuracy and 89.6% adjacent accuracy on 6-level Bloom's classification

CCS Model Validation

Method	Exact Accuracy	Adjacent (±1)	F1 Macro
Random	16.8%	44.5%	0.168
Verb Heuristic	60.6%	74.4%	0.613
TF-IDF + SVM	77.7%	85.3%	0.776
LLM Panel (4-vote)	74.2%	84.3%	0.744
CCS (BERT)	82.0%	89.6%	0.819

Considerations for Using the Data

Ethical Considerations

• This dataset is intended for research in educational AI and benchmark evaluation
• Silver labels have ~82% accuracy; use with appropriate uncertainty quantification
• Questions are sourced from educational materials and do not contain sensitive content

Limitations

• Silver labels may contain systematic biases from the CCS model
• Subject distribution is uneven (STEM-heavy due to OpenStax sources)
• Bloom's Taxonomy classification inherently involves subjectivity

Citation

If you use CogBench in your research, please cite:

@misc{kunuku2026cogbench,
  title={CogBench: A Benchmark for Evaluating Cognitive-Level Control in LLM Question Generation},
  author={Kunuku, Mourya Teja},
  year={2026},
  url={https://huggingface.co/datasets/mouryat9/CogBench}
}

License

CC-BY-4.0