Emergentic AI
Sign in

The dataset atlas

Different windows onto human behavior

Showing 24 of 130 matching entries · 130 total
B01 / Beliefs & preferencesPublic release

Twin-2K-500 follows 2,058 people in the US through several rounds of surveys and decision tasks. Their answers cover preferences and psychology, and some tasks are repeated later.

Where the data comes from
The authors publish the data on Hugging Face under a CC BY 4.0 license.
What we want to find out
If we give a model someone’s earlier survey answers, can it predict that person’s answers to questions we have kept hidden? We want to know whether it captures differences between people, rather than giving everyone the most common answer.
Possible uses
Customer research teams could test whether simulated respondents answer like the people they represent, including the differences between those people. This would help assess their use in customer studies and personalized services.
Source preview ↗
R01 / Related researchPublic release

SocSci210 collects people's responses from social science experiments in a common format. Its paper reports 210 studies, about 400,000 participants and 2.9 million responses. These include responses to different versions of an experiment, allowing us to study how a change in the task affects people's answers.

Where the data comes from
The data is available on Hugging Face, with Socrates code and models in the linked repository. The terms attached to each original source still apply.
What we want to find out
Can a model predict how people respond to a new experiment, including how their answers change when researchers change the task? Such forecasts could help researchers choose which studies to run with people.
Possible uses
Research teams could use the forecasts to choose which customer or product experiments to run with people.
Source preview ↗
C01 / Conversations about placesPublic release

In Talk the Walk, one person plays a tourist trying to find their way through an NYC neighborhood, and another gives directions. They see different information. The dataset records their conversation alongside their actions, maps and views of the streets.

Where the data comes from
The authors provide a public research repository and paper. The project is archived, so we would need to check that the required data and images can still be downloaded.
What we want to find out
A tourist has received directions but has not yet reached the destination. Can we predict where they go next or whether they ask for help?
Possible uses
We could test whether seeing the streets helps an agent predict these decisions better than reading the conversation alone. Before using it to predict failed meetings or deliveries, we would need to check whether the release includes failed as well as successful conversations.
Source preview ↗
B02 / Beliefs & preferencesPublic release

This follow-up to Twin-2K-500 contains 19 studies, their survey instructions, human responses and released model results. Some participants also appear in the original collection, so the two releases are not independent groups of people.

What we want to find out
Can earlier answers about a person help us predict how they respond to a new experiment? We would also ask whether the model predicts the difference between people given different versions of the task.
Source preview ↗
B03 / Beliefs & preferencesPublic release

OpinionQA collects Pew survey questions and the percentage of people choosing each answer. It includes results for the population as a whole and for demographic groups.

Where the data comes from
The authors publish this test of language models in the opinions_qa research repository. Its survey results come from Pew.
What we want to find out
For a new survey question, can we predict the percentage of people who choose each answer? We would check both the population as a whole and differences between demographic groups.
Possible uses
Market researchers and service designers could test whether simulated populations reproduce differences in opinion. Matching a group's percentages would not tell us whether a model can predict any particular person's answer.
Source preview ↗
B04 / Beliefs & preferencesPublic release

SubPOP records the percentages of survey respondents choosing different answers and provides models trained to predict those percentages. It draws on Pew and General Social Survey data also used by other entries in this atlas. Download terms apply.

What we want to find out
After learning from one set of surveys, can a model predict the percentages choosing each answer on questions it has not seen?
Source preview ↗
B05 / Beliefs & preferencesPublic release

Pew's American Trends Panel surveys people over time about their opinions. Its public releases include the questions, answers and survey weights: numbers used to adjust the results so the sample better represents the population.

Where the data comes from
Pew provides successive survey releases through its American Trends Panel dataset pages, subject to its terms of use.
What we want to find out
Can earlier surveys help us predict answers in a later survey? Predicting how one person changes would also require records that let us follow the same person over time.
Possible uses
Researchers could check whether AI-generated survey answers reproduce differences in opinion across groups. They would also need to account for people who stop taking part before drawing conclusions about how opinions change over time.
Source preview ↗
B06 / Beliefs & preferencesPublic release

The General Social Survey asks people about their attitudes, reported behavior and demographic circumstances. Its public data lets us compare answers across groups and survey years. Detailed locations require separate access.

What we want to find out
Can we predict how many people choose each survey answer in a later year, and how those answers differ between groups?
Source preview ↗
B07 / Beliefs & preferencesPublic release

The World Values Survey asks people in several countries about their values and attitudes. Access requires registration, and its terms restrict redistribution.

What we want to find out
If a model learns about values in one country or year, can it predict survey answers in another? This would help us identify where a simulated population gives a misleading picture of people’s views.
Source preview ↗
B08 / Beliefs & preferencesRequest access

The European Social Survey repeatedly surveys attitudes across several countries and documents how the questions are collected. Access could not be confirmed during the earlier review because the portal timed out.

What we want to find out
Can a model predict survey answers across European countries and later survey rounds? We would check whether differences in language or question wording explain its mistakes.
Source preview ↗
B09 / Beliefs & preferencesPublic release

Psych-101 brings together records of the individual choices people make in psychological experiments. Its published collection covers 160 experiments, allowing a model to see earlier trials and try to predict a later response. The records are public.

What we want to find out
After watching someone make several choices in an experiment, can we predict their next choice and, where recorded, how confident they will be? The tasks let us study learning and risk-taking under known rules.
Source preview ↗
B10 / Beliefs & preferencesRequest access

This study builds AI representations of people using interviews, surveys or both, then tests predictions of their responses. The June 2026 revision updates work originally titled Generative Agent Simulations of 1,000 People. Code is public; the repository says the complete interview collection is not publicly downloadable.

What we want to find out
Does an interview help a model predict a person’s later answers more accurately than a survey does? We could also test whether giving it both sources of information helps.
Source preview ↗
B11 / Beliefs & preferencesAccess unconfirmed

Two studies, each with approximately 4,000 participants, examine how people judge their own performance. The paper reports an anonymized data release, but the earlier review could not access its files or check their field definitions because OSF returned an access error.

What we want to find out
Can we predict both someone’s test score and the score they think they will get? Comparing the two could help us understand when people overestimate or underestimate their ability.
Source preview ↗
B12 / Beliefs & preferencesAccess unconfirmed

This 2019 study examines how well people recognize their own performance, an ability researchers call metacognition. It reports released data and analyses, but the OSF download has not been verified.

What we want to find out
When two people perform equally well on a task, do they judge their performance differently? We would test whether a model can predict those self-assessments, including among people with lower scores.
Source preview ↗
B13 / Beliefs & preferencesAccess unconfirmed

This 2022 study examines a proposed explanation of the Dunning–Kruger effect: that people who lack a skill also have difficulty recognizing their mistakes. It reports raw and processed data and analysis code, although the earlier review could not access the OSF files.

What we want to find out
Why do some people misjudge how well they performed? We could compare explanations that blame limited awareness of mistakes with explanations based on noisy scores, then check which predicts new participants more accurately.
Source preview ↗
B14 / Beliefs & preferencesAccess unconfirmed

People answer Cognitive Reflection Test questions, which often invite a tempting wrong answer, and assess their own performance. The study pairs those answers with self-assessments and reports public data.

What we want to find out
Some questions invite a quick but wrong answer. Can we predict who gives that answer and how confident they are, using their recorded responses and self-assessments?
Source preview ↗
C02 / Conversations about placesPublic release

In these recorded conversations, people work out a travel plan, divide resources or resolve a disagreement. Each person knows something the other needs, so they must exchange information to reach an agreement. The conversations and task environments are public.

What we want to find out
Two people have different preferences and must agree on a plan. From what they have told each other so far, can we predict the plan they will accept and whether it meets their needs?
Source preview ↗
C03 / Conversations about placesPublic release

People role-play customers and staff in hotel, restaurant, transport and other service conversations. The public data includes their goals and the service details tracked during the exchange. Releases differ, so choose a specific version.

What we want to find out
A customer asks for a hotel, restaurant or journey with particular requirements. Can we predict what they will book, which alternative they will request, or whether they will leave without completing the booking?
Source preview ↗
C04 / Conversations about placesPublic release

People act out conversations between a traveler and a booking operator. Each task supplies travel requirements and booking information, and the records include the expected booking or cancellation. These are public role-play records rather than real purchases.

What we want to find out
A customer and travel operator are discussing the available flights. Can we predict which booking or cancellation they agree to, while respecting the customer’s travel requirements?
Source preview ↗
C05 / Conversations about placesPublic release

Two people work together on household tasks in a computer simulation. The records show what they say, what they can see and which actions they take. The data is public.

What we want to find out
One person tells another how to complete a household task in a computer simulation. Can we predict when the person carrying out the task will need a clearer instruction or make a mistake?
Source preview ↗
C06 / Conversations about placesPublic release

One person, the architect, describes a structure made of blocks. Another, the builder, constructs it in a shared three-dimensional world. The released data pairs their conversation with a record of the building actions.

What we want to find out
One person describes a structure while another builds it from blocks. Can we predict the builder’s next move and the explanation they will need if they put a block in the wrong place?
Source preview ↗
C07 / Conversations about placesPublic release

IGLU collects instructions for building with blocks in a Minecraft-like computer world, together with actions and views of the scene. Some records contain an ongoing conversation; others contain an instruction collected separately, without someone replying as the task unfolds.

What we want to find out
Suppose someone is told where to place a block. Can a model work out the intended position, or recognize that the instruction leaves more than one possibility? Where the records include a back-and-forth conversation, we could also study how people ask for clearer instructions.
Source preview ↗
C08 / Conversations about placesPublic release

Conference attendees wear sensors that record when they are close to one another. The public records contain the timing of these encounters. They do not record speech or establish that a conversation takes place.

What we want to find out
From an attendee’s earlier encounters and the event schedule, can we predict whom they will stand near later and whether they will meet the same people again? Event organizers could use this to study opportunities for people to meet.
Source preview ↗
C09 / Conversations about placesPublic release

These public sensor records show when office occupants are near one another over time. They let us follow repeated encounters among colleagues, but do not provide their conversations.

What we want to find out
Can we predict when colleagues will be near each other, including colleagues from different teams? These timings could help us make office simulations resemble the opportunities for contact in a real workplace.
Source preview ↗

Dataset / source