Presentation title:
Tower LLM: An Intentional Multilingual Large Language Model for Translation-related Tasks
Presentation description:
In this presentation we will present TowerLLM an LLM that is state of the art for several translation related tasks even when competing with much bigger models. We present a recipe for training such a model and how it involved over time. We describe the importance of accounting and planning for different languages from scratch, how the quality of data and data curation plays a central role during training, and how training for different tasks makes the model better at each task.
We present some real use cases results of tower in production scenarios.