Overview

This repository contains a lightweight library for evaluating language models. We are open sourcing it so we can be transparent about the accuracy numbers we're publishing alongside our latest models.

Benchmark Results

Model	Prompt	MMLU	GPQA	MATH	HumanEval	MGSM1	DROP1 (F1, 3-shot)	SimpleQA
o1				MATH-5002
o1	n/a3	91.8	75.7	96.4	n/a	89.3	90.2	42.6
o1-preview	n/a	90.8	73.3	85.5	92.4	90.8	74.8	42.4
o1-mini	n/a	85.2	60.0	90.0	92.4	89.9	83.9	7.6
GPT-4o
gpt-4o-2024-11-20	assistant	85.7	46.0	68.5	90.2	90.3	81.5	38.8
gpt-4o-2024-08-06	assistant4	88.7	53.1	75.9	90.2	90.0	79.8	40.1
gpt-4o-2024-05-13	assistant	87.2	49.9	76.6	91.0	89.9	83.7	39.0
gpt-4o-mini-2024-07-18	assistant	82.0	40.2	70.2	87.2	87.0	79.7	9.5
GPT-4 Turbo and GPT-4
gpt-4-turbo-2024-04-09	assistant	86.7	49.3	73.4	88.2	89.6	86.0	24.2
gpt-4-0125-preview	assistant	85.4	41.4	64.5	86.6	85.1	81.5	n/a
gpt-4-1106-preview	assistant	84.7	42.5	64.3	83.7	87.1	83.2	n/a
Other Models (Reported)
Claude 3.5 Sonnet	unknown	88.3	59.4	71.1	92.0	91.6	87.1	28.9
Claude 3 Opus	unknown	86.8	50.4	60.1	84.9	90.7	83.1	23.5
Llama 3.1 405b	unknown	88.6	50.7	73.8	89.0	91.6	84.8	n/a
Llama 3.1 70b	unknown	82.0	41.7	68.0	80.5	86.9	79.6	n/a
Llama 3.1 8b	unknown	68.4	30.4	51.9	72.6	68.9	59.5	n/a
Grok 2	unknown	87.5	56.0	76.1	88.4	n/a	n/a	n/a
Grok 2 mini	unknown	86.2	51.0	73.0	85.7	n/a	n/a	n/a
Gemini 1.0 Ultra	unknown	83.7	n/a	53.2	74.4	79.0	82.4	n/a
Gemini 1.5 Pro	unknown	81.9	n/a	58.5	71.9	88.7	78.9	n/a
Gemini 1.5 Flash	unknown	77.9	38.6	40.9	71.5	75.5	78.4	n/a

Background

Evals are sensitive to prompting, and there's significant variation in the formulations used in recent publications and libraries. Some use few-shot prompts or role playing prompts ("You are an expert software programmer..."). These approaches are carryovers from evaluating base models (rather than instruction/chat-tuned models) and from models that were worse at following instructions.

For this library, we are emphasizing the zero-shot, chain-of-thought setting, with simple instructions like "Solve the following multiple choice problem". We believe that this prompting technique is a better reflection of the models' performance in realistic usage.

We will not be actively maintaining this repository and monitoring PRs and Issues. In particular, we're not accepting new evals. Here are the changes we might accept.

• Bug fixes (hopefully not needed!)
• Adding adapters for new models
• Adding new rows to the table below with eval results, given new models and new system prompts.

This repository is NOT intended as a replacement for https://github.com/openai/evals, which is designed to be a comprehensive collection of a large number of evals.

Evals

This repository currently contains the following evals:

• MMLU: Measuring Massive Multitask Language Understanding, reference: https://arxiv.org/abs/2009.03300, https://github.com/hendrycks/test, MIT License
• MATH: Measuring Mathematical Problem Solving With the MATH Dataset, reference: https://arxiv.org/abs/2103.03874, https://github.com/hendrycks/math, MIT License
• GPQA: A Graduate-Level Google-Proof Q&A Benchmark, reference: https://arxiv.org/abs/2311.12022, https://github.com/idavidrein/gpqa/, MIT License
• DROP: A Reading Comprehension Benchmark Requiring Discrete Reasoning Over Paragraphs, reference: https://arxiv.org/abs/1903.00161, https://allenai.org/data/drop, Apache License 2.0
• MGSM: Multilingual Grade School Math Benchmark (MGSM), Language Models are Multilingual Chain-of-Thought Reasoners, reference: https://arxiv.org/abs/2210.03057, https://github.com/google-research/url-nlp, Creative Commons Attribution 4.0 International Public License (CC-BY)
• HumanEval: Evaluating Large Language Models Trained on Code, reference https://arxiv.org/abs/2107.03374, https://github.com/openai/human-eval, MIT License

Samplers

We have implemented sampling interfaces for the following language model APIs:

• OpenAI: https://platform.openai.com/docs/overview
• Claude: https://www.anthropic.com/api

Make sure to set the *_API_KEY environment variables before using these APIs.

Setup

Due to the optional dependencies, we're not providing a unified setup mechanism. Instead, we're providing instructions for each eval and sampler.

For HumanEval (python programming)

git clone https://github.com/openai/human-eval
pip install -e human-eval

For the OpenAI API:

pip install openai

For the Anthropic API:

pip install anthropic

Running the evals

python -m simple-evals.simple_evals --list-models

This will list all the models that you can evaluate.

To run the evaluations, you can use the following command:

python -m simple-evals.simple_evals --model <model_name> --examples <num_examples>

This will launch evaluations through the OpenAI API.

Notes

Legal Stuff

By contributing to evals, you are agreeing to make your evaluation logic and data under the same MIT license as this repository. You must have adequate rights to upload any data used in an eval. OpenAI reserves the right to use this data in future service improvements to our product. Contributions to OpenAI evals will be subject to our usual Usage Policies: https://platform.openai.com/docs/usage-policies.

Footnotes

1. We believe these evals are saturated for our newer models, but are reporting them for completeness. ↩ ↩²

2. For o1 models, we evaluate on MATH-500, which is a newer, IID version of MATH. ↩

3. o1 models do not support using a system prompt. ↩

4. assistant system message in OpenAI API doc: "You are a helpful assistant." . ↩