LKValues: Aligning Large Language Models with Sri Lankan Societal Values

TL;DR

This paper proposes LKValues, a resource suite for aligning large language models with Sri Lankan societal values. It addresses cultural biases in existing models and provides a pipeline for country-specific value alignment. LKValues fine-tuning improves model performance in English and Sinhala.

Why It Matters

The research has practical significance as it enables more culturally sensitive and accurate language models for multilingual societies like Sri Lanka. This can lead to better natural language processing applications and more effective communication. It also highlights the importance of considering local values in AI development.

Key Findings

• Newer and larger LLMs still exhibit cultural value-alignment gaps
• LKValues fine-tuning improves Qwen-family models in English and Sinhala
• Gains from fine-tuning remain model-family dependent