Most of these are based on prompting API-based models. It would be cool if we could dogfood building these components with prompt2model instead of using GPT.
The prompt parser is the obvious first step, as it's quite simple and should be easy for a prompt2model model to do well. Data generation is likely too hard because it requires a generalist model, but something like transforming existing datasets would be a possible target. For autopilot, column selection seems like an easy target.
prompt2model has several components, including:
Most of these are based on prompting API-based models. It would be cool if we could dogfood building these components with prompt2model instead of using GPT.
The prompt parser is the obvious first step, as it's quite simple and should be easy for a prompt2model model to do well. Data generation is likely too hard because it requires a generalist model, but something like transforming existing datasets would be a possible target. For autopilot, column selection seems like an easy target.