How an AI Model Learns the Values of Different Countries: A Dataset of 500,000 Examples

A new dataset for training neural networks has been published: a collection of 500,000 preference examples covering the values of 20 countries, so that language models reflect the worldview of more than just Western users.

Why are AI models skewed toward the West?

Most modern language models are trained mainly on English-language text and calibrated to the preferences of American and European annotators. When such models reason about family roles, attitudes toward authority, religion or individual rights, they systematically underrepresent the value systems of the Global South, Asia and the Middle East.

AI assistants, chatbots and educational systems are used more and more outside the US, and users feel the worldview skew directly. In blind testing by the PLURAL authors, participants from India, Brazil and Japan said that standard LLM responses sound distinctly American.

How the PLURAL dataset works

PLURAL is built on the Integrated Values Survey (IVS), a representative sociological survey covering 92 countries. The researchers developed a two-stage pipeline: real respondents' answers are turned into thematic scenarios, and then each scenario receives a pair of ratings, one more consistent and one less consistent with the values of the given culture.

Key facts about the dataset:

• ~500,000 preference examples in the first public version
• Covers 20 countries from different regions of the world
• Built on the IVS, a survey of 92 countries
• Reduces the mean absolute error of the cultural profile by 27.7% relative to strong baseline models
• 176 blind-test participants from India, Brazil and Japan rated PLURAL responses as more accurate for their own culture

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

Why this matters for AI product developers

Alignment of language models is currently built mainly on RLHF with the preferences of Western annotators. For companies launching AI products in India, Brazil or Arab countries, this means a model may, by default, give answers that are culturally perceived as foreign or inappropriate.

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

What this means

PLURAL is the first large-scale public resource that allows a language model to be trained on the values of a specific country without separate teams of annotators. If the approach proves reproducible, it may 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 model's mean absolute error (MAE) of the cultural profile by 27.7% compared with strong baseline models. This is the result of the automated evaluation described in the preprint.

Which countries took part in the blind testing?

Blind testing was conducted in three countries: India, Brazil and Japan. It involved 176 raters who compared PLURAL responses with standard LLM responses on how well they matched the values of their own national culture.

Why are language models skewed toward the West?

Most modern language models are trained mainly on English-language text and calibrated to the preferences of American and European annotators.