I’m running into a weird bug where Google AI Studio only keeps the Input text but not the Output. I’m wondering if it’s because of the SSML that I enter into Output having extra characters like <> and % (which are required), so it doesn’t recognize the text as valid. It also clears out the output when toggling to other subtabs within Google AI Studio or reload the page, regardless of if I save.
When I refill in the info, it also doesn’t recognize the output when moving that into New Tuned Model, as it only sees the plain text Input. How could this be resolved?
Hi @Logan_Kilpatrick, following up on this. Were you able to find a solution? I was able to temporarily have the SSML output recognized yesterday but it has reverted back to not working again, no does it allow me to train any models with SSML as output.
I have not applied the following answer to SSML specifically. It might not work. But it works in general.
LLMs and that includes the various Gemini are very adept at recognizing patterns. Give them the rules and several examples first and they will apply the pattern. Give them the rules = put the text of the specification defining SSML in the LLM context. Good examples = show in simulated dialog of user and model what model is expected to produce.
@OrangiaNebula As a concept this is true, though in this scenario the issue is not that the LLM can’t create SSML output, but that there is a constraint in Google AI Studio that does not allow for the special characters that SSML requires. If I go into another platform that uses Gemini or similar LLM, it will give me the output as SSML, just not Google AI Studio when typing SSML as output directly within Structured Prompt. The only way around it currently is importing via csv
Thank you, I understand the problem better, AI Studio input field does not handle markup well. I have an idea that might work: try entity-encoding the markup characters. That means you use & l t; for the less-than character and & g t; for the greater-than character. No spaces, I used spaces here because the Discourse UI otherwise converts my entity-encoded characters to the characters they represent.
