cogbench-anonymous/cogbench-passages

CogBench Passages

A curated collection of 120 short academic passages across 8 subjects, used as the source text for the CogBench benchmark — an evaluation of whether LLMs can generate questions that satisfy specified Bloom's-Taxonomy cognitive levels under deterministic, code-checkable constraints.

Anonymized for double-blind review (NeurIPS 2026 Evaluations & Datasets Track). Author and institutional metadata will be added after acceptance.

Evaluative role

This dataset is not a benchmark on its own. It is the source-text input for the CogBench evaluation pipeline, which:

1. Asks an LLM to generate questions about a passage at a target Bloom's level (Remember / Understand / Apply / Analyze / Evaluate / Create).
2. Runs 28 deterministic constraints (no LLM-as-judge) over the question + passage to produce a per-question constraint-satisfaction rate (CSR).
3. Aggregates CSR across passages, subjects, and Bloom's levels.

The dataset is what lets the constraints reference real key_concepts and methods_principles so that constraints like "the question must reference a key concept from the passage" become decidable.

Schema

Field	Type	Description
passage_id	string	Stable identifier, e.g. bio_001, phys_015
subject	string	One of: biology, chemistry, computer_science, economics, history, mathematics, physics, psychology
source	string	OpenStax book slug (e.g. openstax_biology-2e)
section	string	OpenStax section heading (e.g. 1.1 The Science of Biology)
text	string	Passage prose (HTML stripped, equations cleaned). Mean ≈498 words, range 345–782.
key_concepts	list[string]	Auto-extracted top noun-phrases (used by constraints U3, R2, A2)
methods_principles	list[string]	Auto-extracted process/method markers (used by constraint P3)
url	string	Canonical OpenStax URL the passage was extracted from
word_count	int	Whitespace token count

Quick load

from datasets import load_dataset

ds = load_dataset("cogbench-anonymous/cogbench-passages", split="train")
print(ds[0]["passage_id"], ds[0]["subject"], ds[0]["word_count"])

Composition

• Total: 120 passages
• Per subject: 15 each across 8 subjects
• Source: OpenStax open-access textbooks (CC BY 4.0)

Subject	OpenStax book	n
Biology	Biology 2e	15
Chemistry	Chemistry 2e	15
Physics	College Physics 2e	15
Mathematics	Calculus Volume 1	15
Psychology	Psychology 2e	15
Economics	Principles of Economics 3e	15
History	U.S. History	15
Computer Science	Introduction to Computer Science	15

Collection process

• Each subject's canonical OpenStax book was parsed; 15 sections per subject were sampled by length + topical-coverage heuristics.
• Sections shorter than ~200 words or that were exercise/review blocks were dropped.
• key_concepts and methods_principles were extracted with spaCy NER + noun-phrase mining.
• Authors spot-checked passages for math/equation rendering artifacts; passages with unrecoverable LaTeX fragments were dropped.
• Collection window: February–March 2026.

See the full datasheet (Gebru et al. 2021 format) in the source repository: proposal/methods/passages_datasheet.md.

Intended use

• In scope: evaluating Bloom's-Taxonomy-aligned question generation by LLMs.
• Out of scope: reading-comprehension benchmarks, fact-verification corpora, curriculum design, training data for instruction-tuned models.

Limitations

• English-only. Non-English educational content is not represented.
• 120 passages is small; results should be reported with confidence intervals.
• STEM-skewed — 6 of 8 subjects are STEM. Models tuned for STEM may have an unfair aggregate advantage; subject-level results are reported separately.
• OpenStax-derived. Passages are likely present in many LLM pretraining corpora; CogBench reports a contamination probe alongside main results to quantify exposure.
• Deterministic constraints as proxy. The 28 constraints encode common signals of cognitive level (verb choice, named-entity grounding, structural form) but do not directly measure whether a human would judge a question as "Analyze-level". This proxy assumption is part of what CogBench evaluates, not a guarantee.
• Auto-extracted features. key_concepts and methods_principles are heuristic; they reflect what an automatic extractor surfaces, not a curated knowledge graph.

Ethical considerations

• No personal or sensitive data; all content is pedagogical material already on the open web under CC BY 4.0.
• No annotator labels — passages are reused verbatim.

Licensing

Component	License
Passage prose (text, section, url)	CC BY 4.0 (OpenStax)
Auto-extracted key_concepts, methods_principles, schema	Apache 2.0
Build / load scripts	Apache 2.0

Attribution: OpenStax. Textbooks released under Creative Commons Attribution License 4.0. https://openstax.org

Versioning & maintenance

• This is v1.0 (NeurIPS 2026 review).
• A frozen DOI will be minted on Zenodo at camera-ready.
• Future revisions will follow OpenStax book updates (e.g. Biology 2e → 3e); each release will keep prior versions identifiable in the changelog.

Citation

@misc{cogbench2026,
  title  = {CogBench: Deterministic Evaluation of Bloom's-Taxonomy-Aligned Question Generation by LLMs},
  author = {Anonymous},
  year   = {2026},
  note   = {NeurIPS 2026 Evaluations \& Datasets Track submission},
  url    = {https://huggingface.co/datasets/cogbench-anonymous/cogbench-passages}
}

Croissant metadata

The dataset includes a Croissant 1.0 metadata file (croissant.json) with both core and Responsible AI (RAI) fields, as required by the NeurIPS 2026 E&D Track.