How neural networks understand the values of different countries: a dataset of 500,000 examples

A dataset has been published for training neural networks: a collection of 500,000 examples of preferences covering the values of 20 countries, so that language models reflect the worldview of not only Western users.

Why are AI models skewed towards the West?

Most modern language models are trained primarily on English texts and calibrated according to the preferences of American and European annotators. When such models reason about family roles, attitudes towards power, religion, or individual rights, they systematically undervalue the value systems of the Global South, Asia, and the Middle East.

AI assistants, chatbots, and educational systems are increasingly used outside the USA — and the worldview bias is directly felt by users. In a blind test by the PLURAL authors, participants from India, Brazil, and Japan noted that standard LLM answers sound like something American.

How is the PLURAL dataset structured

At the core of PLURAL is the Integrated Values Survey (IVS), a representative sociological survey from 92 countries. Researchers developed a two-stage pipeline: real respondents' answers are transformed into thematic scenarios, and then each scenario receives a pair of ratings — more and less aligned with the values of a given culture.

Key facts about the dataset:

• ~500,000 examples of preferences in the first public version
• Covers 20 countries from different regions of the world
• Built on IVS — a survey from 92 states
• Reduces the average absolute error of cultural profile by 27.7% compared to strong baseline models
• 176 blind test participants from India, Brazil, and Japan recognized PLURAL answers as more accurate for their culture

Authors validated PLURAL using three methods: data-level validation (whether inter-country differences from the original survey are preserved), automatic evaluation of fine-tuning accuracy, and blind testing with live participants in three countries.

Why this matters for AI product developers

The process of aligning language models is currently built primarily through RLHF with preferences of Western annotators. For companies launching AI products in India, Brazil, or Arab countries, this means: the model by default may give answers that are culturally perceived as alien or inappropriate.

The traditional solution — hiring local annotators in each country — is expensive and scales poorly. PLURAL offers an alternative path: use data from representative sociological surveys as a proxy for cultural preferences. The dataset is published in open access on HuggingFace.

What this means

PLURAL is the first large-scale public resource that allows deliberately training a language model to the values of a specific country without engaging separate annotation teams. If the approach proves reproducible, it could change how AI companies localize models for non-Western markets.

Frequently asked questions

How much does PLURAL improve cultural alignment?

Fine-tuning on PLURAL reduces the average absolute error (MAE) of the model's cultural profile by 27.7% compared to strong baseline models — this is the result of automatic evaluation described in the preprint.

Which countries participated in the blind test?

Blind testing was conducted in three countries: India, Brazil, and Japan. It involved 176 raters who compared PLURAL answers with standard LLM answers for alignment with their national culture's values.

Why are language models skewed towards the West?

Most modern language models are trained primarily on English texts and calibrated according to the preferences of American and European annotators.