HighSchoolBench-OEQ

GitHub • Master’s thesis

Dataset description

HighSchoolBench-OEQ contains 2,952 open-ended questions from high-school standardized assessments and science competitions. It is the open-ended component of HighSchoolBench, a multimodal benchmark for evaluating language and vision-language models on high-school assessment questions.

The questions come from assessments such as AP, IB, SAT, PSAT, NAEP, and PISA, U.S. state examinations, and U.S. science olympiads. They include text-based questions and questions involving images, diagrams, and tables.

Each question includes a free-text reference answer and metadata identifying its source assessment, subject, domain, and expected answer type. Answers range from numerical values and mathematical expressions to explanations and worked solutions. The multiple-choice component is available separately as HighSchoolBench-MCQ.

Dataset statistics

The dataset contains 2,952 questions and 531 images.

Split	Questions
train	449
validation	141
test	2,362
Total	2,952

The source split val is exposed as validation on Hugging Face. The split field retains the original source label.

Configurations

The dataset is organized into five configurations according to the supporting context. All configurations contain free-text reference answers.

Configuration	Supporting context
context_text_only__text_response	Text
context_image_only__text_response	Images
context_table_only__text_response	Tables
context_table_image__text_response	Tables and images
context_text_contextualized_image__text_response	Images with associated passages

The context category describes the supporting material accompanying the question; the question stem is stored separately in question.

In the contextualized configuration, individual images are paired with their associated passages. This configuration contains train and test splits; the other configurations also include validation.

Data format

Main fields

Field	Description
id	Question identifier from the source file.
assessment_name	Source assessment or competition.
assessment_type	Assessment stage or exam type.
subject	Subject assigned to the question.
domain	Topic within the subject.
context	Supporting passage or shared context, when available.
question	Question stem. Mathematical expressions use LaTeX.
answer	Free-text reference answer.
split	Original split label: train, val, or test.
answer_type	Expected form of the answer.
context_policy	Whether the question relies on the provided context, domain knowledge, or both.

Context-specific fields

The following fields appear where relevant to the configuration:

Field	Description
image	List of images accompanying the question. A single image is stored as a one-item list.
image_context	Text associated with the images.
table	Table content represented as a Markdown string.
table_context	Text associated with the table.
image_N, image_context_N	Individual image and its associated passage in the contextualized configuration.

Answer types

The answer_type field uses the following categories:

Value	Answer format
numerical_value	Numerical answer
numerical_value_with_units	Numerical answer with units
equation_or_expression	Equation or mathematical expression
short_answer	Short textual answer
brief_explanation	Brief explanation
long_explanation	Extended explanation
calculation_with_work	Calculation with working steps
multi_part_answer	Answer containing multiple parts

The context_policy values are context_only, domain_knowledge_only, and context_first_then_domain_knowledge.

Available files

The dataset is provided in two formats:

• Parquet: data/<configuration>/, with typed columns and embedded images. These files are used by the Hugging Face dataset viewer and load_dataset.
• JSON Lines and images: subsets/<configuration>/<split>/, containing a metadata.jsonl file and the images referenced by its records.

Usage

Load a configuration using the Hugging Face datasets library:

from datasets import load_dataset

dataset = load_dataset(
    "SciKnowOrg/HighSchoolBench-OEQ",
    "context_image_only__text_response",
)

print(dataset)

Access a question and its reference answer:

example = dataset["test"][0]

print("Context:", example["context"])
print("Question:", example["question"])
print("Answer type:", example["answer_type"])
print("Reference answer:", example["answer"])

# Question images are decoded as PIL images.
images = example["image"]

List the available configurations:

from datasets import get_dataset_config_names

print(get_dataset_config_names("SciKnowOrg/HighSchoolBench-OEQ"))

To download the JSON Lines files and accompanying images:

hf download SciKnowOrg/HighSchoolBench-OEQ \
    --repo-type dataset \
    --include "subsets/*" \
    --local-dir HighSchoolBench-OEQ

Citation

If you use HighSchoolBench, please cite:

@misc{rani2026highschoolbench,
  title        = {HighSchoolBench: A Multimodal Benchmark of High School Assessment Questions},
  author       = {Rani, Krishna and D'Souza, Jennifer},
  year         = {2026},
  publisher    = {Hugging Face},
  howpublished = {\url{https://huggingface.co/datasets/SciKnowOrg/HighSchoolBench-MCQ},
                  \url{https://huggingface.co/datasets/SciKnowOrg/HighSchoolBench-OEQ}}
}

HighSchoolBench was introduced in the following master’s thesis:

@mastersthesis{rani2026highschoolbench_thesis,
  title  = {Toward Robust LLM Benchmarking in High School Science Education:
            Constructing a QA Corpus from Standardized Tests},
  author = {Rani, Krishna},
  school = {Gottfried Wilhelm Leibniz Universit{\"a}t Hannover},
  year   = {2026},
  doi    = {10.15488/22429},
  url    = {https://doi.org/10.15488/22429}
}