太长不看。Persona Atlas 能将一位公众人物转化为一幅可实际度量的行为画像。你只需输入一个名字,一个小模型智能体就会在开放网络上研究该人物,撰写一份有据可查的档案,然后以其口吻回答一组固定的开放式“思考”问题。每个回答都会被转化为嵌入向量,这样一来,人物形象就不再是一大段文字,而变成了空间中的一个点。将几位思想家并排放置,你就能看出谁倾向于怀疑,谁倾向于幽默,谁倾向于冷静的抽象。其背后的核心理念是:人格主要关乎风格,而非算力,因此即便进行对话的模型是小模型,这种人格特质也能得以保留。而这,正是本次“小模型构建黑客松”的全部意义所在。
观看简短导览:研究一个人物形象,比较几个形象,阅读特质热力图。
播放器加载失败?在 YouTube 上观看。
引发此项目的那个问题
如果你能让苏格拉底、丘吉尔和一位硅谷创始人共处一室,向他们提出同一个无法回答的问题,然后观察他们各自以多么不同的方式寻求答案,那会怎样?大多数基准测试衡量的是模型知道什么。而 Persona Atlas 追求的是更难捉摸的东西,即某个特定思维是如何运作的,并且它试图让这一点变得可见,而不仅仅是声称如此。
运作方式
一次运行包含三个步骤。
首先,研究。一个具备工具调用能力的智能体执行真实的网络搜索,提取人物画像,并整理出一份公开档案、一份有据可查的事实清单(每条事实都链接回它实际访问过的来源),以及一份“风格假设”,即它对这个人会如何应对一个前所未见的问题所做的最佳猜测。
其次,该人物形象回答基准测试:十个关于身份、伦理、真理、自由意志、意义和机器意识的刻意设计的开放式提示词。这些问题没有正确答案,这是故意的。正是这类问题,会让一个人的人格特质流露出来,而非模型的原始能力。
第三,每个回答都变成一个嵌入向量。这便将每个人物形象转化成了可比较的点:将两个形象并排放置,就能测量它们回答之间的距离。
比较思维
选择任意已保存的角色,对比视图会执行两项操作:它会衡量这些角色在嵌入向量空间中的答案相距多远,给出一个数值来反映整个群体的分歧程度;同时,它会根据十个特质锚点(严谨性、清晰度、创造力、怀疑精神、自信、友善、幽默感、好奇心、务实性、抽象性)对每个角色进行评分,并以特质倾向热力图的形式呈现。
该网格采用双中心化处理,这比听起来要重要得多。一个暖色单元格绝不意味着在某种绝对意义上“该特质得分高”。它表示,与你恰好放在一起的其他角色相比,这个角色更倾向于该特质。将几个截然不同的角色并排放置,各行便会拉开差距:一个角色在幽默感和自信方面呈现暖色,另一个则在抽象性和怀疑精神方面呈现暖色。
底层机制
所有功能均通过 Hugging Face 推理提供商在托管的小型模型上运行:一个紧凑的生成器驱动智能体,一个轻量级嵌入模型负责几何计算,外加实时网络和图像搜索用于事实锚定。前端采用 Gradio,包含三个标签页:研究一次运行、比较已保存的角色、以及检查完整的智能体追踪记录,以便你自行验证它是否依赖真实来源,而非凭空捏造。系统预置了一套角色,因此页面加载后即可进行比较,无需任何 token。
立即尝试
首先打开“比较已保存的角色”标签页,或者研究一个新角色并将其添加到角色图谱中:huggingface.co/spaces/build-small-hackathon/persona-atlas
TL;DR. Persona Atlas turns a public figure into a behavioral portrait you can actually measure. You type a name, a small-model agent researches that person on the open web, writes up a grounded dossier, and then answers a fixed set of open-ended "thinking" questions in their voice. Every answer gets embedded, so a persona stops being a wall of prose and becomes a point in space. Line several thinkers up next to each other and you can see who reaches for skepticism, who for humor, who for cold abstraction. The bet underneath it: personality is mostly style, not horsepower, so it survives even when the models doing the talking are small. Which, this being the build-small hackathon, is sort of the whole point.
Watch the short tour: research a persona, compare a few, read the trait heatmap.
Trouble loading the player? Watch it on YouTube.
The question that started it
What if you could put Socrates, Churchill, and a Silicon Valley founder in the same room, hand them the same unanswerable question, and watch how differently each one reaches for an answer? Most benchmarks measure what a model knows. Persona Atlas is after something harder to pin down, which is how a given mind moves, and it tries to make that visible instead of just claiming it.
How it works
A run has three steps.
First, research. A tool-calling agent runs real web searches, pulls a portrait, and puts together a public profile, a list of grounded facts (each one linked back to a source it actually visited), and a "style hypothesis" that's its best guess at how this person attacks a problem they've never seen before.
Second, the persona answers the benchmark: ten deliberately open-ended prompts about identity, ethics, truth, free will, meaning, and machine consciousness. There are no right answers, on purpose. These are the questions where a personality leaks through instead of the model's raw capability.
Third, every answer becomes an embedding. That turns each persona into points you can compare: put two side by side and measure the distance between their answers.
Comparing minds
Pick any of the saved personas and the comparison view does two things. It measures how far apart their answers sit in embedding space, giving you one number for how much the whole group diverges, and it scores each persona against ten trait anchors (meticulousness, clarity, creativity, skepticism, confidence, kindness, humor, curiosity, pragmatism, abstraction), drawn as a trait-leaning heatmap.
The grid is double-centered, which matters more than it sounds. A warm cell never means "high on this trait" in some absolute sense. It means this persona leans toward that trait more than the others you happened to put on the table. Drop a handful of very different people side by side and the rows pull apart. One runs warm on humor and confidence, another on abstraction and skepticism.
Under the hood
Everything runs on small, hosted models through Hugging Face Inference Providers: a compact generator driving the agent, a lightweight embedding model doing the geometry, plus live web and image search for grounding. The front end is Gradio, with three tabs: research a run, compare saved personas, and inspect the full agent trace, so you can check for yourself that it leans on real sources rather than quietly making things up. A set of personas ships prebuilt, so the comparison works the moment the page loads, no token required.
Try it
Open the Compare saved personas tab to start, or research someone new and add them to the atlas: huggingface.co/spaces/build-small-hackathon/persona-atlas