Verifiers

Will Brown's verifiers grew from a GRPO experiment into a standard library for defining RL environments that LLMs train against, now maintained by Prime Intellect. An environment bundles a dataset, a rollout harness, and a scoring rubric into an installable Python module, and the same object serves three jobs: an eval suite runnable against any OpenAI-compatible endpoint, an agent harness with multi-turn tool use, and a training environment consumed by trainers like prime-rl for large-scale GRPO runs. Environment types span single-turn, multi-turn, and tool-calling agents, with sandboxed execution available for code tasks. Hundreds of community environments published on Prime Intellect's Environments Hub install with one command, which has turned the format into a de facto interchange standard for RL tasks and agent evals. The library itself is MIT licensed, requires Python 3.11 or newer, installs with pip install verifiers, and needs no GPU: evaluation talks to remote or local inference servers, and heavy training is delegated to a separate trainer.