Verifiers

Verifiers is Prime Intellect’s open-source library for building reinforcement learning environments and evaluations for large language models. With roughly 4,700 stars on GitHub and rapid development through 2026, it has become a common foundation for researchers training models with RL on verifiable tasks, from math and code to tool use and agentic workflows.

Overview

RL training is only as good as the environments and reward signals behind it. Verifiers provides a modular toolkit for defining environments (datasets, tasks, interactive settings), attaching programmatic verifiers that score model outputs, and running evaluations or full training loops with popular RL stacks. Its abstractions separate the task definition from the training machinery, so the same environment can serve for benchmarking a frontier model today and fine-tuning your own model tomorrow. Prime Intellect built the library for its own distributed training work, and it shows in the details: environments are versioned, shareable packages, and results are comparable across runs because the scoring code travels with the task definition.

Key Features

• Modular environment definitions for RL training and LLM evaluation
• Programmatic verifiers and rubrics for scoring model outputs reliably
• Support for single-turn, multi-turn and tool-using agentic tasks
• Integrations with popular RL training frameworks and Prime Intellect’s stack
• Reusable datasets and environment packages shared by the community
• Consistent harness for comparing models on identical verifiable tasks
• Designed to pair with distributed training setups, from a single node to large clusters

💳 Pricing

Free and open source under the MIT license. Training runs on your own GPUs or on compute you rent; the library itself costs nothing.

Who Is It For

LLM researchers and post-training teams who need trustworthy environments and reward signals to train or evaluate models on reasoning, coding and agentic tasks.