# ZeroEval

> Open-source evaluation framework for testing large language models with zero-shot prompting on reasoning and coding tasks.

ZeroEval is an open-source evaluation framework designed to benchmark large language models (LLMs) using zero-shot prompting techniques. The project focuses on assessing model capabilities across reasoning, mathematics, and coding tasks without requiring few-shot examples, providing a standardized way to compare different AI models' performance.

The framework evaluates models on multiple benchmark datasets including MMLU-Redux for general knowledge, MATH-500 for mathematical reasoning, CRUX for code understanding, and ZebraLogic for logical reasoning puzzles. ZeroEval maintains public leaderboards that track performance across various model families including OpenAI, Anthropic, Google, Meta, and open-source alternatives.

**Key Features:**

- **Zero-Shot Evaluation** - Tests models without providing example solutions, measuring true generalization capabilities and reasoning abilities across diverse problem types.

- **Multiple Benchmark Support** - Includes MMLU-Redux (knowledge), MATH-500 (mathematics), CRUX (code reasoning), and ZebraLogic (logic puzzles) for comprehensive model assessment.

- **Public Leaderboards** - Maintains transparent rankings of model performance with detailed breakdowns by task category and difficulty level.

- **Open Source Framework** - Fully open-source codebase available on GitHub, allowing researchers and developers to run evaluations locally and contribute improvements.

- **Reproducible Results** - Provides standardized evaluation protocols ensuring consistent and comparable results across different model evaluations.

- **Multi-Model Support** - Compatible with various LLM providers and architectures, enabling fair comparisons between proprietary and open-source models.

To get started with ZeroEval, clone the GitHub repository and follow the installation instructions in the documentation. The framework supports running evaluations through command-line interfaces, making it accessible for researchers conducting model comparisons. Results can be submitted to the public leaderboard for community visibility and benchmarking purposes.

## Features
- Zero-shot LLM evaluation
- MMLU-Redux benchmark
- MATH-500 mathematical reasoning
- CRUX code understanding
- ZebraLogic logical reasoning
- Public leaderboards
- Multi-model support
- Reproducible evaluation protocols
- Open-source framework

## Integrations
OpenAI models, Anthropic Claude, Google Gemini, Meta Llama, Mistral, Qwen, DeepSeek

## Platforms
WEB, API

## Pricing
Open Source

## Links
- Website: https://zeroeval.com
- Documentation: https://docs.zeroeval.com/
- Repository: https://github.com/zeroeval/zeroeval-ts
- EveryDev.ai: https://www.everydev.ai/tools/zeroeval