面向 Agent 的数据

为什么智能体 AI 需要开放数据，以及合成数据如何帮助我们扩展规模。

图：Nemotron 训练后 v3 提示图谱

不仅仅关乎模型权重

构建 AI Agent 很难，因为真实世界并不像基准测试那样运行。

一个无法从失效的 API 调用或从未见过的流程中恢复的 Agent，并不是真正的 Agent。它只是一个带着工具的自动补全器。从后者到前者的转变，是一个数据问题：软件工程轨迹、工具使用失败、多步推理、检索、安全、用户模拟、流程执行，以及最终与物理世界的交互。这些正是 NVIDIA Nemotron 开放数据产品的用武之地。

NVIDIA 近期强调了开放模型如何推动 AI 研究，并指出在知名国际机器学习大会（ICML）上，近 145 篇论文引用了 Nemotron 模型和数据集。合成数据在整个生态系统中扮演着重要角色：

• Nemotron-CC 利用合成数据增强了流行的 Common Crawl 数据集，用于预训练。
• Nemotron-CC-MATH 利用合成数学问题提升推理能力。
• Nemotron Pretraining 是一个覆盖通用、代码、数学和合成数据的广泛集合，包含数万亿 token。

NVIDIA 之所以发布开放数据集，部分原因是为了与社区一起学习，将这些应用拓展到更多场景。

开放权重固然重要。但对于 Agent 而言，权重只是故事的一部分。可复现性同样依赖于模型背后的数据集、策展选择、训练配方和评估方法。

Agent 的行为需要是可检查的。如果模型调用工具、执行流程、检索信息并跨系统行动，开发者就需要理解塑造这些行为的数据。开放数据让 Agent 的行为变得可检查和可解释。而合成数据正是实现这一点的关键拼图。

像对待秘密一样保存它

NVIDIA 应用深度学习研究副总裁 Bryan Catanzaro 最近提到："每家公司都围绕一个秘密建立"——一种竞争对手没有的流程、语料库或客户模式。这些秘密让 AI 变得有用，但公司不应轻易暴露它们。合成数据为团队提供了一种方法，既能保留有用的信号，又无需暴露底层来源。

Bryan 还谈到要培育一个多元且参与式的 AI 生态系统，让各类公司、研究者、政府和社区都能参与贡献。这不仅是一个价值主张，更是一个数据主张。

如果每个模型都从同一狭窄的数据池中学习，当我们发现模型开始变得雷同时，不应感到意外。难点在于，最有用的数据往往掌握在那些不能或不愿直接公开发布的组织手中。更丰富的共享数据层能让每个人受益，但没人愿意第一个拱手让出让自己变得独特的东西。

公开发布的合成数据，是改变这种局面的一个途径。

探索 Agent 数据

作为 Nemotron 开放数据的一部分，我们已发布超过 10 万亿预训练 token 和数百万后训练样本，覆盖多个领域和数据形态。要理清这些数据并非易事——原始的表格数据帮不上什么忙。

为了方便探索 Nemotron 后训练数据中的真实内容，我们构建了 Nemotron 训练后 v3 提示图谱：一个交互式的可视化地图，每个点都代表一个提示样本，这些样本来自 Nemotron v3 后训练集合，并通过按量采样反映数据混合的真实比例。

颜色叠加和过滤器可以让你按数据集、管线阶段、领域或工具使用情况重新组织图谱。由于语义相似的提示会聚集在一起，你可以放大到某个区域——编码算法、安全、数学、智能体行为——查看代表性示例，并利用这些信号来策展数据、构建评估，或理解模型为何表现出特定行为。

人物角色万岁

Agent 还需要理解它们所要服务的人，这里“数据质量”变得本地化，而非普适。一个在英文互联网数据上训练的毒性分类器，可能会漏掉韩语或日语中的敌意信息，因为在那些语言中，敌意常常编码在敬语层级中，而非明显的词汇里。相同的信号，不同的语境。团队已经在用这种方式为 Agent 打下本地化基础。

Nemotron-Personas 是解决这一问题的一种尝试：基于本地情况的合成人物角色，捕捉人群的多样性和复杂性。使用 NVIDIA 最先进的、用于生成合成数据的复合 AI 工具 NeMo Data Designer 构建，Nemotron-Personas 映射了官方的地区人口和地理统计数据。其目标不是复现真实的人，而是一种帮助开发者测试其系统是否反映了它们声称服务的用户、语言、地区和职业的方式。上个月在巴黎的 VivaTech 大会上，我们的系列中推出了第十个国家的人物角色，目前已覆盖超过 24 亿人口。

当质量是本地化的时候，只有了解当地情况的人才能构建它——区域研究者、母语者、领域专家，以及能够与你一起检查和纠正的利益相关者。这就是公开学习：不是孤立地发布数据，而是协同构建。

真实基准

合成数据需要作为数据源系统的一部分进行整合。这里存在权衡。它可以降低风险，但并不能取代对真实基础、数据血统、策展、评估和人类判断的需求。

一种有用的思考方式是借助"合成阈值"：即那些不再能被视为纯真实数据的临界点。这条界线并不总是显而易见。真实工作流、人类反馈、模型生成的轨迹、模拟用户和合成标签都可能交织在一起。答案不是假装合成数据是虚假的或无害的，而是要记录下哪些是生成的，哪些有真实依据，哪些经过了审查，以及数据旨在测试什么。随着越来越多的 AI 系统在人工信息上训练，我们需要更好的共享习惯来检查、记录并公开讨论这些技术。

质量在不同语境下也有着不同的含义。推理数据需要更难的问题和更干净的轨迹。人物角色数据需要分布忠实度和本地审查。智能体流程需要任务多样性、故障覆盖和恢复路径。这个领域目前仍然更多是一门手艺，而非公式。

这就是开放方法之所以重要的原因。合成数据不仅仅是生成更多样例的问题。它关乎提出更好的问题，并使那些否则无法坐到同一张桌前的各方成为可能：公司不必泄露其秘密，政府不必损害隐私，研究者不必等待可能永远不会来临的批准。

AI 领域稀缺的资源不是 token，而是组织间的信任。合成数据是我们构建这种信任的少数工具之一。

Data for Agents

Why agentic AI needs open data, and why synthetic data is how we scale it.

Image: Nemotron Post-Training v3 Prompt Atlas

More Than Model Weights

Building AI agents is hard, because the real world does not behave like a benchmark.

An agent that can't recover from a broken API call, or a workflow it has never seen, is not really an agent. It is an autocompleter with tools. Getting from one to the other is a data problem: software engineering traces, tool-use failures, multi-step reasoning, retrieval, safety, user simulation, workflow execution, and eventually physical world interaction. That is where NVIDIA Nemotron's open data products live.

NVIDIA recently highlighted how open models are driving AI research and showing up across the popular International Conference on Machine Learning (ICML), with nearly 145 papers citing Nemotron models and datasets. Synthetic data plays an important role across that ecosystem:

• Nemotron-CC used synthetics to enhance the popular Common Crawl dataset for pretraining.
• Nemotron-CC-MATH leverages synthetic math questions to improve reasoning.
• Nemotron Pretraining is a broad collection spanning general, code, math, and synthetic data across trillions of tokens.

Part of why NVIDIA releases open datasets is to learn with the community to expand upon these various applications.

Open weights matter. But for agents, weights are only part of the story. Reproducibility also depends on the datasets, curation choices, training recipes, and evaluation methods behind the model.

Agent behavior needs to be inspectable. If a model calls tools, executes workflows, retrieves information, and acts across systems, developers need to understand the data that shaped those behaviors. Open data makes agent behavior inspectable and explainable. Synthetic data is a key piece of the puzzle to making that possible.

Keep It Like a Secret

NVIDIA's VP of Applied Deep Learning Research Bryan Catanzaro recently noted: "every company is built around a secret" — a workflow, corpus, or customer pattern competitors don't have. Those secrets make AI useful, but companies shouldn't casually expose them. Synthetic data gives teams a way to preserve useful signals without exposing the underlying sources.

Bryan also talks about cultivating a diverse and participatory AI ecosystem where many kinds of companies, researchers, governments, and communities can contribute. That is not just a value claim. It is a data claim.

If every model learns from the same narrow pool of data, we should not be surprised when the models start to feel the same. The hard part is that the most useful data often sits inside organizations that cannot or will not publish it directly. Everyone benefits from a richer shared data layer. No one wants to be the first to give away the thing that makes them special.

Synthetic data, released openly, is one way to change that math.

Exploring Agent Data

As part of Nemotron open data, we've released over 10 trillion pre-training tokens and millions of post-training samples spanning many domains and data shapes. That's a lot to make sense of — and raw dataset tables don't help much.

To make it easier to explore what's actually in Nemotron post-training data, we built the Nemotron Post-Training v3 Prompt Atlas: an interactive visual map where each point is a prompt sample, drawn from the Nemotron v3 post-training collection and volume-sampled to reflect the honest proportions of the data mixture.

Color overlays and filters let you reorganize the map by dataset, pipeline stage, domain, or tool use. Since semantically similar prompts cluster together, you can zoom into a region — coding algorithms, safety, math, agentic behavior — inspect representative examples, and use that signal to curate data, build evals, or understand why a model behaves the way it does.

Viva La Persona

Agents also need to understand people they are built to support, and this is where “data quality” becomes local, not universal. A toxicity classifier trained on English internet data can miss hostile messages in Korean or Japanese, where aggression is often encoded in politeness levels rather than obvious vocabulary. Same signal, different context. Teams are already grounding agents this way.

Nemotron-Personas is one attempt at addressing that: locally grounded synthetic personas capturing the diversity and complexity of populations. Built using NeMo Data Designer, NVIDIA’s state-of-the-art compound-AI tooling for synthetic data generation, Nemotron-Personas mirrors official regional demographic and geographic statistics The goal is not to recreate real people. In a way, it’s to help developers test whether their systems reflect the users, languages, regions, and occupations they claim to serve. Last month at VivaTech in Paris, we launched our tenth country in the collection, which now represents more than 2.4B people.

When quality is local, only people who know that locality can build it — regional researchers, native speakers, subject-matter experts, stakeholders who can inspect and correct alongside you. That's learning in public: not releasing data in isolation, but building it collaboratively.

Ground Truths

Synthetic data needs to be integrated as part of a system of data sources. There are tradeoffs. It can reduce risk, but it does not remove the need for grounding, lineage, curation, evaluation, and human judgment.

One useful way to think about this is with "synthetic thresholds": points where data can no longer be treated as purely real. That line is not always obvious. Real workflows, human feedback, model-generated traces, simulated users, and synthetic labels can all become intertwined. The answer is not to pretend synthetic data is fake or harmless. It is to document what was generated, what was grounded, what was reviewed, and what the data is meant to test. As more AI systems are trained on artificial information, we need better shared habits for inspecting it, documenting it, and debating these technologies in public.

Quality also means different things in different contexts. Reasoning data needs harder problems and cleaner traces. Persona data needs distributional fidelity and local review. Agentic workflows need task diversity, failure coverage, and recovery paths. The field is still more craft than formula.

That is why open methods matter. Synthetic data is not just about generating more examples. It is about asking better questions, and making it possible for parties who otherwise could not sit at the same table: companies without giving away their secrets, governments without compromising privacy, and researchers without waiting for permission that may never come.

The scarce resource in AI is not tokens. It is trust between organizations. Synthetic data is one of the few tools we have for building it.

We hosted a livestream on Tuesday, July 7, 2026 on Why Open Data Matters with an amazing panel. It is worth checking out, along with the Nemotron data collections on Hugging Face.

Stay up to date on NVIDIA Nemotron by subscribing to NVIDIA news and following NVIDIA AI on LinkedIn, X, YouTube, and the Nemotron channel on Discord.

Access open Nemotron Models on Hugging Face and a collection of NIM microservices and Developer Examples on build.nvidia.com.