Loading…
2
Serious question, I trained a tiny model on 400 of my grandpa's letters and it nailed his voice on the first try
My grandpa passed in 2019 and I've had a shoebox of about 400 letters he wrote to my grandma sitting in a closet in Dayton. Two weekends ago I typed 60 of them into a plain text file, fed them into a small open model on my laptop with a cheap tutorial I found, and asked it to finish a sentence he never wrote. It came back with "the corn's coming in slow but the dog don't mind" and I just sat there, because that is exactly how he talked. What surprised me is the model only needed 60 letters, not all 400, and the trick was keeping his spelling mistakes in instead of cleaning them up. I learned that small and messy beats big and polished for this kind of thing. Has anyone else messed with tiny models on old family writing, and did you strip the typos out or leave them in?
1 comments
Log in to join the discussion
Log In1 Comment
clairen8514d ago
Keep the typos. Clean ones lose the person.
0