Scientific Reasoning Benchmark · openai
frontierscience
A benchmark assessing AI models on expert-level reasoning tasks across the natural sciences.
What frontierscience is
FrontierScience is an evaluation benchmark developed by OpenAI to test the capabilities and boundaries of artificial intelligence systems on expert-level tasks across the natural sciences. It targets advanced reasoning to help researchers understand what automated models can and cannot achieve at the frontier of scientific research.
The dataset provides standardized text-based evaluation material designed to reveal specific performance thresholds in high-level scientific problem-solving. It serves researchers who need rigorous evaluation standards when probing model competence on specialist scientific queries.
How it was made
To safeguard against contamination during model pre-training, the records include a canary GUID string that enables automated filters to exclude the benchmark questions from training corpora.
Key facts
- Benchmark category
- Scientific Reasoning Benchmark
- Modality
- Text
- Data format
- JSON
What you can use it for
- Benchmarking expert-level scientific reasoning
- Probing machine learning models on complex reasoning problems drawn from across the natural sciences.
- Mapping capabilities at the research frontier
- Systematically determining which scientific research tasks current AI systems can perform and where they fail.
- Auditing training data contamination
- Using the embedded canary GUID to detect whether evaluation material has inadvertently entered training pipelines.
Explore the data
The schema, real rows, full-text search and a SQL filter — read in your browser, nothing downloaded.
Explore
Read the real rows without downloading anything
Reading rows…
LiveRead from openai/frontierscience at the moment you asked. Nothing is cached or stored — every row above came from that request.
How the columns are distributed
Over 160 rows of default/test
answer
string_text- min
- 1
- median
- 134.50
- mean
- 1471.29
- max
- 21,458
problem
string_text- min
- 408
- median
- 1242.50
- mean
- 1509.76
- max
- 6,589
subject
string_label- physics44%
- chemistry38%
- biology19%
task_group_id
string_text- min
- 36
- median
- 36
- mean
- 36
- max
- 36
Strengths and limits
Good at
- Targets expert-level reasoning across the natural sciences rather than introductory or general-knowledge tasks.
- Incorporates a canary GUID string to help prevent benchmark contamination in language model training corpora.
Watch out for
- Limited to text-only evaluations, excluding non-textual experimental data or multimodal inputs.
Questions
Can I look inside frontierscience without downloading it?
Yes. This page reads its schema, real rows and per-column statistics in the browser, with full-text search and a SQL filter. Nothing is downloaded to your machine.
What licence is frontierscience under?
Its authors declare apache-2.0. Read the terms on the original page before reusing the data.
What scientific fields does the benchmark address?
It targets expert-level reasoning across the natural sciences.
What data formats and modalities are supported?
The dataset is provided entirely as text in JSON format.
Why is there a canary GUID in the dataset?
The canary GUID is included so web scrapers and pipeline developers can prevent the evaluation questions from being incorporated into model training sets.
Written from frontierscience's own dataset card. Statements citing a figure that does not appear there are removed before publication. Licence and running details come from our registry.