Hi, sorry to be such a spammer the past few days: I got my ner.llm.correct working with some examples and spacy-llm with flaky, slow GPT4. But I notice something odd in the labeling ui vs. the LLM response. In short, the response from the LLM has a correct reply for this case:
Seared Scallops with Ginger, Persimmon and
Heirloom Tomatoes | True | DISH | a full dish, with food ingredients and flavors
There is a newline in the middle of it, which the LLM can handle. I have set "allow-newline-highlight" to true in the config so i can correct these things, but it would be nice if Prodigy actually selected the right text span from the model?
Updated: OTOH, this response from the LLM is wrong, but again Prodigy isn't highlighting all the spans. Not sure what's going on... related to the newlines no doubt.
Thanks for the background and the example. Yes, it seems like it could be an issue with the newlines in the llm-io interface. We'll make a ticket and investigate it. We have a few teammates out the rest of the week but we'll respond back when we have any update. Thank you!
Apologies for taking so long on following up on this! I've been testing your example with latest Prodigy version and it seems that after upgrading spacy-llm (currently at 0.7.1 in Prodigy 1.15.3) the issue is gone.
I know it's been a while so I won't be asking to confirm, but if you're still running into these issues after upgrading to latest Prodigy do ping us!
Thanks and, again, apologies for the late reply!