Verifiers: Environments for LLM Reinforcement Learning and Evaluation

Introduction

Verifiers is a powerful Python library developed by Prime Intellect AI, specifically designed for creating robust environments to train and evaluate Large Language Models (LLMs). It provides a comprehensive framework where environments encapsulate everything needed to run and assess a model on a particular task. Each environment typically includes a dataset of task inputs, a harness for the model (managing tools, sandboxes, and context), and a reward function or rubric to score the model's performance. Verifiers is deeply integrated with the Prime Intellect Environments Hub, their prime-rl training framework, and their Hosted Training platform, offering a complete ecosystem for LLM development.

Installation

Getting started with Verifiers is straightforward. First, ensure you have uv and the prime CLI tool installed.

# install uv
curl -LsSf https://astral.sh/uv/install.sh | sh
# install the prime CLI
uv tool install prime
# log in to the Prime Intellect platform
prime login

To set up a new workspace for environment development, use:

# ~/dev/my-lab
prime lab setup

Alternatively, to add Verifiers to an existing project, run:

uv add verifiers && prime lab setup --skip-install

Examples

Verifiers allows you to easily initialize new environment templates. To create a fresh environment module, use the prime env init command:

prime env init my-env # creates a new template in ./environments/my_env

Environment modules are self-contained Python files that expose a load_environment function. Here's a basic example:

# my_env.py
import verifiers as vf

def load_environment(dataset_name: str = 'gsm8k') -> vf.Environment:
    dataset = vf.load_example_dataset(dataset_name) # 'question'
    async def correct_answer(completion, answer) -> float:
        completion_ans = completion[-1]['content']
        return 1.0 if completion_ans == answer else 0.0
    rubric = vf.Rubric(funcs=[correct_answer])
    env = vf.SingleTurnEnv(dataset=dataset, rubric=rubric)
    return env

You can also install environments from the Environments Hub:

prime env install primeintellect/math-python

To run a local evaluation with any OpenAI-compatible model:

prime eval run my-env -m gpt-5-nano # run and save eval results locally

Why Use Verifiers

Verifiers offers a robust solution for anyone working with LLMs, providing a structured and efficient way to:

• Create Custom Environments: Define specific tasks with tailored datasets, model interaction harnesses, and precise reward functions.
• Facilitate Reinforcement Learning: Design environments optimized for training LLMs using reinforcement learning techniques.
• Evaluate LLM Capabilities: Conduct thorough evaluations of model performance across various tasks and metrics.
• Generate Synthetic Data: Leverage environments to produce high-quality synthetic data for further model training or analysis.
• Seamless Integration: Benefit from tight integration with Prime Intellect's broader ecosystem, including their Environments Hub, prime-rl training framework, and Hosted Training platform.
• Streamlined Workflow: The prime CLI tool simplifies environment setup, installation, evaluation, and publishing.

Links

• GitHub Repository
• Official Documentation
• Environments Hub
• PRIME-RL Training Framework