LKValues: Aligning Large Language Models with Sri Lankan Societal Values

ACL ARR 2026 January Submission9278 Authors

Keywords: Bilingual Evaluation, Bilingual Instruction dataset, Cross-lingual Robustness, Cultural Alignment, Pluralist Values, Sinhala, Sri Lanka, Sri Lankan societal values, Survey-grounded value elicitation, Value Alignment, value-sensitive benchmark

Abstract: As Large Language Models (LLMs) increasingly shape educational, civic, and communicative decisions in a society, their value assumptions often default to Western norms inadequately capturing native viewpoints in linguistically diverse settings like Sri Lanka. Current benchmarks rarely capture Sri Lankan values in Sinhala language, limiting culturally grounded evaluation and fine-tuning. To bridge this gap, we propose LKValues, a 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, we derive 40 majority-endorsed societal values. Using these values, we construct LKvaluesIT, a bilingual news-derived instruction corpus containing 300k scenario-based instances, and LKvaluesBench, a value-sensitive evaluation benchmark of 2,000 instances. We Supervised Finetune two open-source models (Qwen3-4B-Instruct, Gemma3-1B-Instruct) and evaluate under various settings. We also evaluate a set of LLMs with LKvaluesBench. Frontier baselines achieve high accuracy, but smaller models show large English-Sinhala gaps. LKValues fine-tuning boosts accuracy in both languages and narrows disparities, with Qwen3-4B-Instruct-LKV gaining 2.5% overall micro-accuracy and, a 13.6% Sinhala increase (vs. 8.7% English decrease). These gains highlight LKValues' efficacy in embedding Sri Lankan values, offering a replicable pipeline for low-resource, country-specific pluralist value alignment. Data and codes will be publicly released.

Paper Type: Long

Research Area: Resources and Evaluation

Research Area Keywords: benchmarking, cross-lingual transfer, multilingualism, less-resourced languages, values and culture, model bias/fairness evaluation, human-centered evaluation, evaluation methodologies, prompting, reproducibility

Contribution Types: Data resources, Surveys

Languages Studied: English, Sinhala

Submission Number: 9278