LKValues: Aligning Large Language Models with Sri Lankan Societal Values

2026-07-22 · Source: Computation and Language · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Data Science & Analytics · Depth: Expert, quick

Summary

LKValues is a newly proposed survey-grounded resource suite designed to align Large Language Models (LLMs) with Sri Lankan societal values, addressing existing cultural biases towards Western norms. Developed from a trilingual survey of 205 respondents, it identifies 40 majority-endorsed values. The suite comprises LKvaluesIT, a 150,000-instance Sinhala-English news-derived instruction corpus, and LKvaluesBench, a 1,000-instance value-sensitive evaluation benchmark. Initial evaluations of proprietary and open-weight LLMs, including Qwen3.5-4B-Base, Qwen3.5-9B-Base, and Aya-Expanse-8B-Base, revealed persistent low-resource and cultural value-alignment gaps. Fine-tuning experiments demonstrated that LKValues significantly improves Qwen-family models in both English and Sinhala, reducing invalid outputs and cross-lingual disparities, thereby proving its efficacy in embedding specific cultural values. The dataset is publicly available.

Key takeaway

For NLP Engineers developing LLMs for diverse global markets, this research highlights the necessity of country-specific value alignment. You should integrate survey-grounded resources like LKValues to mitigate Western cultural biases and improve model performance in low-resource languages. Consider adapting this replicable pipeline to create localized instruction corpora and evaluation benchmarks for your target regions. This approach ensures more culturally sensitive and accurate LLM outputs, reducing invalid responses and cross-lingual disparities.

Key insights

LLMs can be culturally aligned with specific societal values using survey-grounded, country-specific datasets and fine-tuning.

Principles

• LLM value alignment often exhibits Western bias.
• Survey-grounded data improves cultural alignment.
• Fine-tuning reduces cross-lingual disparities.

Method

A trilingual survey identifies majority-endorsed values. These inform an instruction corpus and an evaluation benchmark. LLMs are then fine-tuned and evaluated using these resources.

In practice

• Create survey-grounded value datasets.
• Fine-tune LLMs for specific cultural contexts.
• Evaluate LLMs with country-specific benchmarks.

Topics

• LLM Value Alignment
• Cultural Bias
• Sri Lankan Values
• Sinhala Language
• Instruction Tuning
• Evaluation Benchmarks

Code references

• NextME14/LKValues

Best for: Research Scientist, AI Scientist, NLP Engineer, AI Ethicist

Related on AIssential

• LKValues: Aligning Large Language Models with Sri Lankan Societal Values — Culturally biased LLMs necessitate country- and language-specific alignment resources to handle local societal values effectively.
• IndoBias: A Dual Track Culturally Grounded Benchmark for LLMs Bias Evaluation in Indonesian Languages — Culturally-grounded benchmarks are crucial for evaluating LLM bias in diverse, multilingual contexts like Indonesia.
• AlignCultura: Towards Culturally Aligned Large Language Models? — Cultural alignment in LLMs requires systematic evaluation against UNESCO principles to prevent biased outputs.
• EvalMORAAL: Interpretable Chain-of-Thought and LLM-as-Judge Evaluation for Moral Alignment in Large Language Models — EvalMORAAL offers a transparent CoT and LLM-as-judge framework to evaluate moral alignment, revealing significant regional disparities.
• Jul 13, 2026Societal ImpactsClaude’s values across models and languages — Claude's expressed values can be quantified into four axes, revealing variations across models and languages, impacting user experience.
• Cross-Lingual Bias in Large Language Models: A Comparative Analysis of English and Swahili — Social biases in LLMs transform, not transfer, across languages, making English-only audits insufficient for multilingual deployment.

See Counsel's argued verdicts on the open AI decisions leaders are weighing →

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Computation and Language.