LKValues: Aligning LLMs with Sri Lankan Societal Values in Sinhala and English

Forum topic · 小凯 · 2026-07-24

Summary

A forum post introducing LKValues (arXiv:2507.18391), the first survey-grounded resource suite for aligning large language models with Sri Lankan societal values. Value alignment in LLMs remains culturally biased toward Western norms, leaving multilingual societies like Sri Lanka underserved. From a trilingual survey of 205 respondents combining adapted global frameworks with LLM-elicited local constructs, the researchers derived 40 majority-endorsed societal values. These were used to build LKvaluesIT, a 150k-instance Sinhala-English news-derived instruction corpus, and LKvaluesBench, a 1,000-instance value-sensitive evaluation benchmark. Evaluations of proprietary and open-weight LLMs, plus fine-tuning of Qwen3.5-4B-Base, Qwen3.5-9B-Base, and Aya-Expanse-8B-Base, show that even newer and larger LLMs exhibit low-resource and cultural value alignment gaps. LKValues fine-tuning improved Qwen-family performance in English and Sinhala, reducing invalid outputs and cross-lingual disparities, though gains depend on the model family. The datasets are publicly available on GitHub, providing a replicable pipeline for low-resource, country-specific pluralistic value alignment.

Overview

Research area: NLP Authors: Nethmi Muthugala, Supryadi, Surangika Ranathunga Published: 2026-07-24 arXiv: 2507.18391

Abstract (translated summary)

Value alignment of Large Language Models (LLMs) has been shown to be culturally biased toward Western norms. This results in the mishandling of local values in multilingual societies such as Sri Lanka that have their unique cultural dynamics. Existing benchmarks overlook Sri Lankan-contextualized values in its official language Sinhala, hindering culturally sensitive evaluation and fine-tuning.

To bridge this gap, the authors propose LKValues, the first survey-grounded resource suite for Sri Lankan value alignment. From a trilingual survey of 205 respondents, blending adapted global frameworks and LLM-elicited local constructs, they derive 40 majority-endorsed societal values. Using these values, they construct:

LKvaluesIT — a Sinhala-English news-derived instruction corpus containing 150k scenario-based instances
LKvaluesBench — a 1,000-instance value-sensitive evaluation benchmark

They use LKvaluesBench to evaluate a set of proprietary and open-weight LLMs, and fine-tune three open-weight base models (Qwen3.5-4B-Base, Qwen3.5-9B-Base, and Aya-Expanse-8B-Base).

Findings

Newer and larger LLMs still exhibit low-resource and cultural value alignment gaps.
LKValues fine-tuning improves Qwen-family model performance in both English and Sinhala, reducing invalid outputs and cross-lingual disparities.
Benefits remain dependent on the model family.

The results highlight LKValues' effectiveness in embedding Sri Lankan values and provide a replicable pipeline for low-resource, country-specific pluralistic value alignment.

Resources

Datasets are publicly available at: https://github.com/NextME14/LKValues

---

*Auto-collected on 2026-07-24.*