From Fortune.: 2024-06-06 11:18:58

Unbabel, a tech company specializing in translation services, unveiled a new AI model, TowerLLM, outperforming OpenAI’s GPT-4o in English to six languages. The model excelled in translating between English and various languages, surpassing GPT-4o and GPT-4 by narrow margins in most cases. TowerLLM also showed expertise in professional domains.
Unbabel’s TowerLLM model’s success in translation challenges OpenAI’s GPT-4 performance, especially with low resource languages. The model surpasses GPT-4o and GPT-4, improving translation accuracy in specific fields like finance, medicine, law, and technical writing by 1-2% compared to OpenAI’s models. TowerLLM’s multilingual capabilities stand out for diverse reasoning tasks as well.
TowerLLM’s training methodology sets it apart, emphasizing multilingual text datasets and careful curation for better quality translations. Unbabel utilized COMETKiwi for assessing translation quality and finetuning TowerLLM with a high-quality dataset. Coined as one of many smaller models outperforming larger ones, TowerLLM showcases the power of quality data over model size in AI evolution.



Read more at Fortune.: Unbabel says new TowerLLM AI model beats OpenAI’s GPT-4 at translation