Train with yaml?
Yaml has better clarity, easier to read by human. And maybe LLM too?
After that we can manually convert them to json with online tools or with other LLM
Yaml has better clarity, easier to read by human. And maybe LLM too?
After that we can manually convert them to json with online tools or with other LLM
Greetings. I believe this is a really good idea. It’s definitely worth trying. I’ll attempt to create a small test dataset of cards in YAML format and train a model. If it does turn out to be convenient, I think I’ll create a proper YAML dataset and focus on that.
Currently, I’m working on a large number of mappings between natural text --- semi-structured text --- and the strict JSON format used in Silly Tavern, including converting existing cards between the formats. In theory, YAML should fit perfectly as a middle ground between natural text and JSON. I suspect there might not even be a need for additional tools to convert YAML to JSON — CardProjector should handle format conversions directly, though it will require some time to refine the dataset.
Thank you for the idea; I will definitely work on it and run a couple of experiments.