Each line in your input JSONL file should represent a single, self-contained prediction request with its corresponding prompt and any necessary schema information directly applicable to that specific request.

If your three individual requests are truly distinct in their purpose, prompts, and desired output schemas, it might be more appropriate to run three separate batch prediction jobs. Each job would then use its own input JSONL file, tailored to a specific prompt and expected output schema.

Feel free to browse the best practices for batch predictions.