kesimeg/Turkish-synthetic-personas

Turkish-synthetic-personas

Dataset Overview

This dataset is an open source synthetically generated persona dataset. To generate these personas a pipeline similar to Nemotron Persona generation pipeline was used. The dataset is grounded with real world demographic distribution of Turkiye using different statistical information provided by Turkish Statistical Institute (TÜİK). The grounding data includes city, age, gender, education, employment status and marital status.

The aim of this data is to be used in other data generation pipelines to increase diversity in the dataset. For example, if one wants to generate an instruction-following dataset the personas here can be used to increase the diversity of users.

The Probabilistic Graphical Model (PGM) part in the pipeline was built using TÜİK data. Sector data was randomly assigned to each user with equal probability. The rest of the fields in each persona was generated using Apache-2.0 licensed google/gemma-4-26B-A4B-it model.

How to use it

You can load the dataset with the following lines of code.

from datasets import load_dataset

nemotron_personas = load_dataset("kesimeg/Turkish-synthetic-personas")

Limitations

The goal here is not replicating the distribution of Turkiye one-to-one. For example due to low diversity in occupation "Sector" attribute was added by randomly sampling from a list of different sectors. The sector information might not fit well with the other attributes of the persona such as city, education etc.

Additionally AI models are known to be biased as well. For example the number of doctors is pretty low for a population of this size. Another example is that around 40% of the population has the hobby of "doğa yürüyüşleri" which is an unlikely number for a population.

Be aware of these limitations while using the dataset!

Capturing Turkish Culture

Some personas have "Okey" or "Tavla" as hobbies which indicates that the model captured the Turkish culture up to some degree. While further analysis is needed to verify this it shows signs of cultural alignment.

License/Terms of Use

This dataset is licensed under the Creative Commons Attribution 4.0 International License (CC BY 4.0). You can freely use this dataset for any commercial or non-commercial project.

Intended use

The intended use of this dataset is for increasing diversity of synthetically generated datasets by incorporating personas in a generation pipeline and using its diversity.

Dataset Details

The dataset contains:

• —17k personas across (10 persona fields & 14 contextual fields)
• —6 grounding contextual fields

Dataset Description & Quality Assessment

You can find some statistics below, regarding the dataset.

Age & Gender Distribution

Age data was generated using population counts reported by TÜİK. The seed data reported the number of people in an age group based on their city and gender. The generated data is parallel with the original. <center> <img src="images/gender_dist.png" width="600px"> </center>

Education Level by Age Group

The heatmap shows the education level vs age. You can see that younger generations are much more likely to have a university degree which can be explained by higher number of universities compared to past. <center> <img src="images/educationheatmapmanual.png" width="600px"> </center>

Population Distribution by City

The figure below shows the Top-10 cities with the highest population which is in line with the real life data. <center> <img src="images/top10cities.png" width="600px"> </center>

Cross-Persona diversity

Each persona type (arts persona, sports persona etc.) is based on a domain. As you can see in the figure below they form clusters based on the domain meaning the LLM was able to generate distinct personas traits for each domain. <center> <img src="images/personadistby_types.png" width="600px"> </center>

Persona population diversity

Each persona type (arts persona, sports persona etc.) is like a dimension of the same persona. The seed data to generate those personas are the same. So all those dimensions together compose a single persona. After combining those persona dimensions into a unified text a single persona is formed. Below you can see the distribution of those personas. <center> <img src="images/population_dist.png" width="600px"> </center>

Data Collection Method

• —Hybrid: Human, Synthetic

Citation

If you find the data useful, please cite:

@misc{kesim2026turkishpersonas,
  author = {Kesim, Ege},
  title = {kesimeg/Turkish-synthetic-personas},
  month = {September},
  year = {2026},
  url = {https://huggingface.co/datasets/kesimeg/Turkish-synthetic-personas}
}