LLMs improve Greek reasoning with SFT and RL, bypassing accuracy metrics

A new research paper explores the capabilities of large language models when fine-tuned for low-resource languages, using Greek as a case study. The study found that while standard accuracy benchmarks showed minimal improvement, supervised fine-tuning (SFT) significantly enhanced the models' ability to reason and generate responses in Greek, with improved grammaticality and reduced token usage. However, SFT also introduced new defects, such as format inconsistencies and a tendency to revert to English. Reinforcement learning with verifiable rewards proved effective in fixing these SFT-induced issues and further improved the models' adherence to instructions. AI

IMPACT Demonstrates that SFT and RL can significantly improve LLM performance in low-resource languages, even when accuracy metrics fail to capture gains.

RANK_REASON Research paper detailing model capabilities and training methods. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv stat.ML →

• Alibaba Group
• Christos Petrocheilos
• Greek
• NVIDIA
• OpenAI

• paper
• model release

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →