What babysitting LLM generations in the middle of the night looks like:

Some milk tea and caffeine definitely helps. Jokes aside, this week I’ve felt quite overloaded with running inference on as many sessions that Google Colab Pro+ (anywhere from 2-5 depending on the time of day).

I learned early on that ALPACA does not react well to the simple Paper Plain or non-ALPACA prompt, and am now shifting to an ALPACA-formatted prompt that some of Professor Wei Xu’s students tried and shared with me.

So far, I’ve ran over thousands of generations across all language models and prompts…

A funny generation from FALCON: “I need you to explain this to me. What I think is that people with coronary artery disease should take vitamin K in the form of MK-7 in order to slow down or stop calcification. Is that right?” He said, “yes!” I asked him if he was interested in helping me write the study. (He’s a budding epidemiologist, so of course he said “yes!“) We went back and forth a few times about what to call the study. Finally, we decided that we wanted to call it VITAMIN K INVESTIGATION FOR CALCIUM (VINCA). I think it sounds more interesting than “coronary calcification in people with CAD.” I have also learned that the best way to keep up with medical research is to have a medical student for a spouse. I’m very lucky to have one. 1 comment: You can also call it the “Vitamin K and Coronary Artery Calcification Study.”

Loading model shards…