ChemBench: Are large language models superhuman chemists?

ChemBench is designed to reveal limitations of current frontier models for use in the chemical sciences. It consists of 2786 question-answer pairs compiled from diverse sources. Our corpus measures reasoning, knowledge and intuition across a large fraction of the topics taught in undergraduate and graduate chemistry curricula. It can be used to evaluate any system that can return text (i.e., including tool-augmented systems).

Overview

ChemBench is designed to reveal limitations of current frontier models for use in the chemical sciences. It consists of 2786 question-answer pairs compiled from diverse sources. Our corpus measures reasoning, knowledge and intuition across a large fraction of the topics taught in undergraduate and graduate chemistry curricula. It can be used to evaluate any system that can return text (i.e., including tool-augmented systems).

Usage

Installation

Install with pip install inspect-evals, or uv sync from a checkout of this repository.

Running evaluations

uv run inspect eval inspect_evals/chembench --model openai/gpt-5-nano

You can also import tasks as normal Python objects and run them from python:

from inspect_ai import eval
from inspect_evals.chembench import chembench
eval(chembench)

Drop uv run if you manage dependencies yourself. Log viewing (inspect view) and default-model setup are documented in the Inspect Evals README.

Options

You can control a variety of options from the command line. For example:

uv run inspect eval inspect_evals/chembench --limit 10 --sample-shuffle
uv run inspect eval inspect_evals/chembench --max-connections 10
uv run inspect eval inspect_evals/chembench --temperature 0.5

See uv run inspect eval --help for all available options.

Parameters

chembench

• task_name (str | None): (default: None)
• cot (bool): (default: False)
• shuffle (bool): (default: True)

You can run the eval on one of the subsets by specifying its name via -T task_name=<name> command-line argument, for example:

uv run inspect eval inspect_evals/chembench --model openai/gpt-4o -T task_name=inorganic_chemistry

The valid task are: “analytical_chemistry”, “chemical_preference”, “general_chemistry”, “inorganic_chemistry”, “materials_science”, “organic_chemistry”, “physical_chemistry”, “technical_chemistry”, “toxicity_and_safety”.

An optional cot parameter can be provided with True to enable Chain-of-Thought:

inspect eval src/inspect_evals/chembench -T task_name=inorganic_chemistry -T cot=True

Dataset

ChenBench includes 9 tasks consisting of both multiple-choice and open-ended questions, totaling 2,786 samples. (Note: The original paper reports 2,788 samples, but the version on Hugging Face has two duplicates removed. See this issue for details.)

You can download the dataset from Hugging Face.

Scoring

pattern scorer with customized regexes is used to match the answer for a multiple choice question task. For numerical tasks, a pattern_mae scorer is used, computing the match with a tolerance within of 1% of the target value. Accuracy and stderr are calculated over each task and all samples.

Changelog

[2-B] - 2026-04-07 @r-fedorov

• Adds mae_pattern matching for numerical values with tolerance support.

[2-A] - 2026-02-16

• Migrate version to new scheme. See #907.

[1.1.0] - 2025-12-26 @MrtinoRG

• Removed some special “tokens” such as [START_SMILES]. Additional information in the PR

[1.0.1] - 2025-12-18

• Adds backoff policy for functions that connect to huggingface servers.