Prime Intellect 发布 Verifiers v1：用于智能体强化学习训练与评估的可组合任务集、Harness 与运行时

Prime Intellect 发布了 0.2.0 验证器：https://www.primeintellect.ai/blog/verifiers-v1。它预览了一个重写的核心，发布时使用新的 verifiers.v1：https：//www.primeintellect.ai/blog/verifiers-v1 命名空间。现代评估现在运行编码代理和工具、压缩和子代理。因此，v1 重建了环境以大规模运行这些代理工作负载。

首先，考虑什么是验证者：Prime Intellect用于智能强化学习和评估的环境栈。此前，环境会将其数据、代理逻辑和基础设施捆绑在一起。相比之下，v1 将该捆绑拆分为三个可组合的部分。

任务集定义了工作内容：数据、工具和评分。安全带解决了任务并产生了滚动。这个线束可以是ReAct环路、CLI代理，或者你自己的。部署随后在运行时内运行，运行时可以是本地的，也可以是在沙盒中。由于这些部件是解耦的，任何任务集都运行在任何兼容的束带下。

定义完这些部件后，接下来的问题是它们如何交流。核心部分是由验证者管理的拦截服务器。它位于代理运行时和推理服务器之间。具体来说，它代理推理的请求和响应。同时，它记录轨迹，设置采样参数，并可重写工具响应。这种重写有助于减少培训中的奖励技巧。

为了规模化，每台服务器会多路复用固定数量的部署，默认为32次。池随后随着观察到的并发性进行弹性扩展。服务器还拥有一个客户端来转发这些请求。在评估过程中，评估客户端作为盲HTTP代理。在培训过程中，TrainClient 会包裹渲染器，实现忠实的 RL 令牌训练。

由于线束使用不同的方言，验证者目前支持三种。这些包括OpenAI聊天完成、OpenAI回复和拟人性消息。方言适配器将每种线格式归一化为规范的vf类型。因此，你的评分逻辑与被测试的药剂保持独立。

架构明确后，考虑团队如何使用它。比如，你可以在Terminal-Bench 2的Codex下运行Nemotron 3 Ultra。

同样，团队可以重复使用 Harbor 数据集而无需重写奖励逻辑。Prime Intellect 仅用一个小类就将 Terminal Bench 2 移植到 v1。在内部测试中，验证器在相同任务上的表现与 Harbor 相匹配。Harbor 是第一个获得全面支持的第三方格式；NeMo Gym 和 OpenEnv 仅处于 Alpha 支持阶段。

在训练方面，相同的环境可以直接接入 prime-rl。在一次长度惩罚消融实验中，GLM-4.5-Air 在六个 H200 节点上使用 ScaleSWE 进行了训练。该运行用时两天，并在 SWE-Bench-Verified 上进行了评估，显示了稳定的智能体训练表现。

每次运行都从定义数据和评分的任务集开始，与任何运行框架无关：

然后，任何任务集都可以通过 TOML 和 CLI 在选定的运行框架下运行：

需要与我们合作推广您的 GitHub 仓库或 Hugging Face 页面或产品发布或网络研讨会等吗？请联系我们：https://forms.gle/wbash1wF6efRj8G58

Michal Sutter 是一名数据科学专业人士，拥有帕多瓦大学的数据科学理学硕士学位。凭借在统计分析、机器学习和数据工程方面的扎实基础，Michal 擅长将复杂的数据集转化为可操作的洞见。

构建一个智能事件场地运营系统 [完整代码]：https://pxllnk.co/twdn5

谢谢！我们的团队将很快与您联系

Prime Intellect launched verifiers 0.2.0 ：https://www.primeintellect.ai/blog/verifiers-v1 . It previews a rewritten core, shipped under the new verifiers.v1：https://www.primeintellect.ai/blog/verifiers-v1 namespace. Modern evaluations now run coding agents with tools, compaction, and subagents. Accordingly, v1 rebuilds environments to run these agentic workloads at scale.

First, consider what verifiers is: Prime Intellect’s environment stack for agentic reinforcement learning and evaluations. Previously, an environment bundled its data, agent logic, and infrastructure together. In contrast, v1 breaks that bundle into three composable pieces.

A taskset defines the work: the data, tools, and scoring. A harness solves the task and produces a rollout. That harness can be a ReAct loop, a CLI agent, or your own. The rollout then runs inside a runtime , either local or in a sandbox. Because the pieces decouple, any taskset runs under any compatible harness.

With those pieces defined, the next question is how they communicate. The central piece is the verifiers-managed interception server . It sits between the agent’s runtime and the inference server. Specifically, it proxies requests to, and responses from, inference. Meanwhile, it records the trace, sets sampling parameters, and can rewrite tool responses. That rewriting helps mitigate reward hacks during training.

For scale, each server multiplexes a constant number of rollouts, defaulting to 32. A pool then scales elastically with observed concurrency. The server also owns a client that relays those requests. During evaluation, an EvalClient acts as a blind HTTP proxy. During training, a TrainClient wraps renderers for faithful token-in RL training.

Because harnesses speak different dialects , verifiers supports three as of now. These are OpenAI Chat Completions, OpenAI Responses, and Anthropic Messages. A dialect adapter normalizes each wire format into canonical vf.types . Consequently, your scoring logic stays independent of the agent tested.

With the architecture clear, consider how teams use it. For example, you can run Nemotron 3 Ultra on Terminal-Bench 2 under Codex.

Similarly, teams can reuse Harbor datasets without rewriting reward logic. Prime Intellect ported Terminal Bench 2 into v1 with only a small class. In its internal testing, verifiers matched Harbor’s performance on the same tasks. Harbor is the first fully-supported third-party format; NeMo Gym and OpenEnv have alpha support.

On the training side, the same environments plug into prime-rl directly. In a length-penalty ablation, GLM-4.5-Air trained on ScaleSWE across six H200 nodes. That run took two days and evaluated on SWE-Bench-Verified, showing stable agentic training.

Each run starts from a taskset that defines data and scoring, independent of any harness:

Any taskset then runs under a chosen harness via TOML and the CLI:

Need to partner with us for promoting your GitHub Repo OR Hugging Face Page OR Product Release OR Webinar etc.? Connect with us ：https://forms.gle/wbash1wF6efRj8G58

Michal Sutter is a data science professional with a Master of Science in Data Science from the University of Padova. With a solid foundation in statistical analysis, machine learning, and data engineering, Michal excels at transforming complex datasets into actionable insights.

Build an Agentic Event Venue Operator [Full Codes]：https://pxllnk.co/twdn5

Thanks! Our team will contact you soon