Just reporting that the Databricks Instruct model does great on structured output. Out of 1223 Job samples, DevOps Engineer and AI Engineer job postings from Indeed, it was able to successfully output the data structure in the screenshot.
These two combined runs we a whopping 1,569,391 Input Toknes and 827997 Output tokens. I was testing extracting a label for the skills, the skill names, and the reference from within the original job description where the skill was taken.
I'm going to go through this to get an idea of the actual output quality, more on that later.
Just wanted to share what I was up to this weekend.
Oh yeah, i logged everything into a local instance of Langfuse. Thinking of dropping the Postgres DB backup for folks to consume, would that be helpful? It's only like 10MB and you can restore it into a Docker Postgres instance.