PrimeIntellect's Verifiers Lets Anyone Build Open RL Training Environments for LLMs
PrimeIntellect's open-source Verifiers library turns RL environments into shareable Python packages, powering everything from quick evals to 100B+ model training runs
Takeaways
- Verifiers is PrimeIntellect's open-source library for building RL training and eval environments for LLMs, now at v0.1.14 with 4,200+ stars.
- Each environment is a self-contained Python package bundling a dataset, a model harness, and a reward function -- shareable via the Environments Hub.
- The latest release adds the v1 Taskset/Harness API, a unified config shape for evals and RL training, and GEPA prompt optimization artifacts.
- GEPA (integrated prompt optimizer) matches or beats GRPO on several tasks using up to 35x fewer rollouts.
- The stack powered INTELLECT-3, a 106B MoE model trained on 512 H200 GPUs that achieves SOTA on math, code, and reasoning benchmarks.
- The entire stack -- Verifiers, prime-rl, and all training environments -- is open-sourced under MIT/Apache 2.0 licenses.
Verifiers is PrimeIntellect's open-source Python library for building, sharing, and running reinforcement learning environments for language models. With over 4,200 GitHub stars and 560 forks, it has quietly become the backbone of a growing open ecosystem for LLM post-training, and the same stack that powered INTELLECT-3, a 106B parameter model that achieved state-of-the-art results on math, code, and reasoning benchmarks.
The problem it's solving
Training LLMs with RL requires more than a dataset. You need a harness (the scaffolding that runs the model, manages tools, handles multi-turn conversations), a reward function, and a way to version and share all of it. Until now, most teams built this from scratch and tightly coupled it to their specific training framework, making it nearly impossible to reuse or share.
This coupling complicates versioning, makes it difficult to run controlled ablations across environment variants, and creates friction for external contributors who must navigate the full training codebase to add or modify tasks or run offline evaluations. Verifiers addresses this by making each environment a self-contained, versioned Python package.
What an environment actually is
Verifiers is a library for creating environments to train and evaluate LLMs. Environments contain everything required to run and evaluate a model on a particular task, and can be used for training models with RL, evaluating capabilities, generating synthetic data, experimenting with agent harnesses, and more.
Concretely, every environment bundles three things:
- Dataset -- the task inputs, which can be static or dynamically generated
- Harness -- tools, sandboxes, context management, and multi-turn interaction logic
- Rubric -- the reward function that scores the model's output
Environments built with Verifiers are standalone Python packages that can be installed and used in your own projects, or shared with the community through the Environments Hub. The Environments Hub at