A company is developing a new foundation model (FM) for multilingual text generation.

Continued pre-training extends the training of an existing foundation model by exposing it to additional datasets that improve its knowledge and capabilities.

For multilingual text generation, a diverse dataset containing multiple languages and dialects helps the model learn different linguistic structures, expressions, and communication patterns.

Including representative language variations also supports responsible AI objectives by reducing the risk that particular linguistic communities are underrepresented.

Amazon Bedrock supports continued pre-training for compatible foundation models using additional unlabeled training data.

Prompt management alone does not improve the model ' s underlying multilingual knowledge. Training separate models and combining them introduces unnecessary complexity.

Model distillation from a single-language teacher may also restrict linguistic diversity.

Therefore, continued pre-training on a representative multilingual dataset is the most appropriate strategy, although fairness and performance must still be evaluated separately.