Multilingual Arbitrage Optimizing Data Pools To Accelerate Multilingual Progress 2024 08 28

Choosing one model to generate training examples for every language can overlook other models' strengths. Research published by Cohere explores selecting among several models according to their performance in different languages. This approach, called multilingual arbitrage, changes which model supplies synthetic examples. The authors report better results than relying on their strongest single teacher, with the largest gains among the least-resourced languages they tested. The useful finding is that the choice of data-generating model should account for language-specific performance. Publisher abstract.