Truth be told, this last week wasn't super interesting. We are kind of in a lull at the moment due to waiting for the funds to roll in from investors, and the CEO is closing some deals, so we're kind of just feeling around for our plans for the future once he can become more involved again.
Last week started off with my generation of 3 different conversations that had a varying amount of pre-included example questions that the bot would draw "inspiration" from. One of the quirks of these LLM systems is that they tend to take the path of least resistance and will start quoting examples verbatim instead of just using them as a springboard, and it's very hard to circumvent this behavior without actually just getting in and fine-tuning it.
So my goal was to seeing how it responded to having zero example questions vs. the original 40 that the CEO had proposed vs. the 10 I'd whittled it down to. I think it gave a good insight as to how much freedom we want the bot to have. It seemed like it was going off on tangents more when there were no examples, but 40 felt too rigid. 10 is probably the sweet spot, even though it still struggles to come up with questions outside of the 10. Perhaps I could whittle them down even further.
Either way, beyond that, I spent my time learning more about machine learning. There's a good course by freeCodeCamp and likewise 3blue1brown has been releasing a series on Deep Learning revolving specifically around LLMs and it's very intriguing seeing all the math play out. I learned a lot from that. I'm also figuring out the best approach to how we want to feed in the fine-tuning data.
Thursday I met up with all three of the psychologists to discuss architecture again and also bring the newest one into the loop as to what we're going to be working on. I told him specifically that we want to curate some training data for conversations for the AI and his help as a specialist in what we're doing is perfect for that.
I read some articles on prompt engineering that were pretty useful and I incorporated some of the tips into the chat prompt and it worked better.
The next day I set out to work on making a prompt to generate a bunch of example conversations to work off of, kind of like I'd done at Lawgoat, but this time for the purpose of imbuing the language we desire. It's going to be a slog but it's important.
And that's pretty much all that happened last week. Till next time, cheers.
Top comments (0)