Training infrastructure for a text-to-text generation system

Design the data and training workflow for a hypothetical conversational text service. Separate pretraining from adaptation, training records from retrieval indexes, and approximate compute work from elapsed training time.

Text-to-text LLMs take text as input and produce text as output. Conversational use adds requirements for following instructions and responding in the context of earlier turns. A model that predicts text is not automatically a reliable assistant; its training, application context, and evaluation all affect the resulting conversation.

The example below illustrates a follow-up question. Answering “What about Japan?” requires resolving it against the preceding question about a capital city. The service must supply that context to the model rather than expecting an isolated request to contain everything needed.