LKValues: Aligning Large Language Models with Sri Lankan Societal Values

2013-08-25 · Source: cs.CL updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Data Science & Analytics · Depth: Expert, extended

Summary

LKValues introduces the first survey-grounded resource suite designed to align Large Language Models with Sri Lankan societal values, addressing the cultural bias of existing LLMs towards Western norms. This initiative developed 40 majority-endorsed values from a trilingual survey of 205 respondents, combining global frameworks with LLM-elicited local constructs. The suite includes LKvaluesIT, a 150k instance Sinhala-English news-derived instruction corpus, and LKvaluesBench, a 1,000-instance value-sensitive evaluation benchmark. Experiments with proprietary and open-weight LLMs, including Qwen3.5-4B-Base, Qwen3.5-9B-Base, and Aya-Expanse-8B-Base, revealed that even newer, larger models exhibit low-resource and cultural value-alignment gaps. Fine-tuning with LKValues significantly improved Qwen-family models in both English and Sinhala, reducing invalid outputs and cross-lingual disparities, though gains varied by model family. The dataset and methodology offer a replicable pipeline for country-specific pluralist value alignment.

Key takeaway

For AI Scientists and ML Engineers deploying LLMs in culturally diverse, low-resource regions like Sri Lanka, recognize that general multilingual models often fail to capture local societal values. You should prioritize implementing country-specific alignment pipelines, leveraging survey-driven value identification and bilingual instruction datasets like LKValues. This approach significantly improves model behavior, reduces invalid outputs, and narrows cross-lingual disparities, ensuring culturally sensitive and reliable LLM deployment.

Key insights

Culturally biased LLMs necessitate country- and language-specific alignment resources to handle local societal values effectively.

Principles

• Value alignment is not culturally uniform.
• Survey-driven methods elicit local societal values.
• Low-resource value alignment needs culturally grounded supervision.

Method

A survey-driven, human-guided pipeline combines international frameworks with LLM-assisted elicitation to identify 40 societal values, then curates bilingual instruction and benchmark datasets.

In practice

• Conduct trilingual surveys for local value identification.
• Curate news-derived instruction datasets for SFT.
• Mix value-aligned data with general instruction data.

Topics

• Large Language Models
• Value Alignment
• Sri Lankan Culture
• Sinhala NLP
• Multilingual LLMs
• Instruction Tuning
• Evaluation Benchmarks

Code references

• NextME14/LKValues

Best for: Research Scientist, AI Scientist, Machine Learning Engineer, NLP Engineer

Related on AIssential

• LKValues: Aligning Large Language Models with Sri Lankan Societal Values — LLMs can be culturally aligned with specific societal values using survey-grounded, country-specific datasets and fine-tuning.
• Opportunities and Challenges of Large Language Models for Low-Resource Languages in Humanities Research — LLMs offer transformative potential for low-resource language research in humanities, despite significant data and ethical challenges.
• PLURAL: A Global Dataset for Value Alignment — PLURAL is a dataset enabling LLMs to align with diverse global values, validated by automated and human evaluations.
• Large language models often prioritize Western moral values, overlooking other cultures — Large language models exhibit a systematic Western moral bias, misrepresenting non-Western cultural values.
• Cross-Lingual Consensus: Aligning Multilingual Cultural Knowledge via Multilingual Self-Consistency — LLMs can self-align cultural knowledge across languages by identifying consistent responses and transferring them to weaker language contexts.
• CuMA: Aligning LLMs with Sparse Cultural Values via Demographic-Aware Mixture of Adapters — CuMA uses demographic-aware routing and specialized adapters to prevent LLMs from collapsing diverse cultural values into a generic average.

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 cs.CL updates on arXiv.org.