Reproduce “Fine-Tuning LLMs for Low-Resource Dialect Translation: The Case of Lebanese” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Silvana Yakhni, Ali Chehab
Researchers at the American University of Beirut demonstrate that fine-tuning large language models on culturally authentic data outperforms training on larger non-native datasets for Lebanese dialect translation, while introducing LebEval - a new culturally-aware evaluation benchmark for assessing translation quality.
Key insights
- Models trained on smaller but culturally authentic datasets consistently outperform those trained on larger non-native datasets
- Contrastive fine-tuning with both correct and incorrect translation examples improves performance more than basic instruction tuning
- Curriculum learning strategies showed limited gains likely due to catastrophic forgetting
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.