My friend and I both trained GPT2 on our chat logs. It's mostly just hilarious seeing what comes out of it, but I've actually gotten real insight out of "hearing myself talk" -- it's similar _enough_ to my personality that it shows me my interests, bad habits etc. And we can ask each other questions, or write the first half of an answer and see what comes out. It can be pretty weird, but we've actually gotten some great advice out of it too. (When you train it on your own text, it still keeps its "wisdom" from the original model.)
If anyone wants to try, I used this colab thing (I don't even need a GPU! Blows my mind that this is free)
I tried asking this in the Show HN thread on that exact colab project, but how difficult would it be to set it up in your own local Jupyter notebook if you're okay using your own GPU?
Edit: Ah, I see in another thread (https://news.ycombinator.com/item?id=22129978) that your GPU needs 11gb+ of VRAM to train the data, which my 1080 certainly doesn't have. A friend of mine works at https://spell.run which offers free trials for anyone interested in an alternative to Google. I may give it a shot this weekend.
My friend said he got it running on 8GB VRAM. But the first time he ran it, I think it wasn't even using his GPU (it took days instead of hours to train though).
If anyone wants to try, I used this colab thing (I don't even need a GPU! Blows my mind that this is free)
https://colab.research.google.com/drive/1VLG8e7YSEwypxU-noRN...
If you use Colab it uploads your data to Google's servers. In this case, they already had our chats (WhatsApp backup to Drive).