How to Generate Retrieval Training and End-to-End Agent-Tuning for a Subset of APIs?

[ChesterXi]

Hello, I'm currently working with a subset of APIs and I'm looking to generate Retrieval Training and perform End-to-End Agent-Tuning for these specific APIs. I would appreciate guidance on the best practices and steps to achieve this. Any relevant documentation or examples would be very helpful.

[Reason-Wang]

Hi, currently our data is mainly converted from ToolBench. If you want to generate some data for your own APIs, I would suggest you to follow the process in their repo, which means you could first sample some APIs, and then generate queries based on sampled APIs using few-shot prompt. This data can be used in retrieval training. Finally, use GPT-3.5 or GPT-4 to inference on these queries and take generated trajectories for agent-tuning.