Verifiers Documentation — RL Environments and Evals for LLM Agents
by Prime Intellect
Write one environment spec that serves both your evals and your RL training run — then publish it as a wheel.
Overview
Verifiers is Prime Intellect's MIT-licensed library for defining the tasks an LLM agent is trained and evaluated on, and the Environments Hub is the public registry those definitions are published to. The v1 documentation runs Overview, Getting Started, Architecture, Tasksets, Evaluation, Harnesses, Env, Harbor and GEPA, with a preserved legacy v0 set — Environments, Evaluation, Training, Development, Reference and FAQs — for code still importing verifiers as vf. The Architecture page is the core of it. A Taskset is the client-owned collection of tasks, loaded once and distributed to workers. A Harness runs inside the rollout runtime and never calls a model provider directly, routing model traffic through an interception server over a local connection or Prime Tunnel. A Trace is assembled live from those intercepted request and response pairs. A Rollout is the executable combination of one loaded task, the harness and any tools, each with independent trace and runtime state. A server-backed orchestrator spawns workers that execute rollouts under a chosen runtime: subprocess for debugging, docker locally, or the prime and modal sandbox runtimes for production and training. Interception also lets you set sampling parameters and rewrite tool responses and search results to block reward hacking. A parallel tutorial track covers the Hub itself: environments are ordinary Python packages declaring dependencies in pyproject.toml, shipped as wheels, installed with the prime CLI, and consumed unchanged by prime-rl for FSDP training or by hosted evaluations. The library was created by Will Brown and has 4.6k stars.
At a Glance
- Topic
- ML
- Level
- Advanced
- Format
- Documentation
- Cost
- Free
- Duration
- Self-paced; ~4-6 hours to read the v1 docs and publish a first environment, plus GPU time for any training run
- Provider
- Prime Intellect
- Hands-on
- Yes — code/exercises
- Certificate
- None
What You’ll Learn
- ✓Define a taskset that an evaluation and an RL run can both consume unchanged
- ✓Separate harness logic from provider calls using the interception server
- ✓Choose a rollout runtime: subprocess for debugging, Docker locally, or remote sandboxes
- ✓Package an environment as a Python wheel and publish it to the Environments Hub
- ✓Read live traces to diagnose where an agent's reward actually comes from
- ✓Block reward hacking by rewriting tool responses and search results at interception
- ✓Install a community environment with the prime CLI and benchmark a model against it
Highlights
- •One spec serves both evaluation and RL training — the same environment feeds prime-rl with no code modification
- •Environments ship as ordinary Python wheels, so sharing one is a package publish rather than a fork of somebody's training stack
- •Reward-hacking countermeasures live in the interception layer instead of being left to each environment author
- •The repo ships an AGENTS.md and a skills/ directory written for coding agents reading it, alongside human-written architecture guides
- •An independent Hugging Face write-up reproduces the whole loop: prime env install primeintellect/alphabet-sort@0.1.5, vf-eval against gpt-4.1-mini (0.982 mean reward) and Qwen3-0.6B (0.403), then roughly eight hours of GRPO on two A6000s lifting Qwen3-0.6B to 0.578
Who It’s For
Best For
- ✓ML engineers doing reinforcement-learning post-training on agentic tasks
- ✓Teams that need reproducible agent evals instead of ad-hoc scoring scripts
- ✓Researchers publishing an environment they want other people to train on
- ✓Anyone migrating off an environment implementation welded to one trainer
Prerequisites
- •Solid Python and comfort with uv-managed projects
- •Working knowledge of RL post-training concepts such as rollouts, rewards and GRPO
- •GPU access, rented or local, for anything beyond evaluation
FAQ
What is Verifiers Documentation — RL Environments and Evals for LLM Agents?
The official documentation for Verifiers, Prime Intellect's MIT-licensed library for defining the tasks an LLM agent is evaluated and trained on, plus the tutorial track for the Environments Hub those definitions are published to. It is for ML engineers doing agentic RL post-training who are tired of environments welded to one training stack.
Is Verifiers Documentation — RL Environments and Evals for LLM Agents free?
Verifiers Documentation — RL Environments and Evals for LLM Agents is free to access.
What level is Verifiers Documentation — RL Environments and Evals for LLM Agents for?
Verifiers Documentation — RL Environments and Evals for LLM Agents is aimed at a advanced audience. Recommended background: Solid Python and comfort with uv-managed projects, Working knowledge of RL post-training concepts such as rollouts, rewards and GRPO, GPU access, rented or local, for anything beyond evaluation.
How long does Verifiers Documentation — RL Environments and Evals for LLM Agents take?
Expect roughly Self-paced; ~4-6 hours to read the v1 docs and publish a first environment, plus GPU time for any training run. Most learners work through it at their own pace.
What will I learn from Verifiers Documentation — RL Environments and Evals for LLM Agents?
You'll learn: Define a taskset that an evaluation and an RL run can both consume unchanged; Separate harness logic from provider calls using the interception server; Choose a rollout runtime: subprocess for debugging, Docker locally, or remote sandboxes; Package an environment as a Python wheel and publish it to the Environments Hub; Read live traces to diagnose where an agent's reward actually comes from; Block reward hacking by rewriting tool responses and search results at interception; Install a community environment with the prime CLI and benchmark a model against it.
Topics
Sources
This page was written from 5 sources, 2 on domains other than docs.primeintellect.ai.