verifiers: Build Environments for LLM Training and Evaluation

Overview

verifiers provides a library for creating environments to train and evaluate LLMs. It gives developers a way to define and work with the environments used in those workflows, rather than treating training and evaluation as separate tasks.

The project is designed to connect with Prime Intellect’s Environments Hub, the prime-rl training framework, and its hosted training platform. It is most relevant to teams building LLM reinforcement-learning environments or using those Prime Intellect services.

Key Features

• Create environments for LLM training and evaluation.
• Use the same library across training and evaluation workflows.
• Integrate with the Prime Intellect Environments Hub.
• Work alongside the prime-rl training framework.
• Connect to Prime Intellect’s hosted training platform.
• Consult human-written architecture guides in the repository’s docs/ directory.

Use Cases

• An LLM researcher can build an environment for training a model with reinforcement learning.
• An evaluation team can create environments to assess LLM behavior.
• A developer publishing or discovering environments through Prime Intellect’s Environments Hub can use the library’s integrated workflow.
• A team already using prime-rl or Prime Intellect Hosted Training can use verifiers to define its environments.

Getting Started

Install the Prime CLI with uv:

uv tool install prime

See the README for setup details and the repository’s docs/ directory for architecture guides.

Alternatives

• llmgym: LLMGym offers interactive environments for developing and evaluating LLM applications, while verifiers focuses on environments for LLM training and reinforcement learning.

Considerations

The project is active, with frequent commits and releases, and has contributions from many developers rather than depending mainly on one. However, the measured issue data shows little recent closure activity, so teams should assess whether their open issues or support needs are being addressed. The repository specifies Python 3.11 to 3.14; it does not list Docker or Compose as requirements. The README does not specify hardware requirements, so check the needs of your intended training workload separately. The MIT license permits broad use, subject to its license terms.