Tried fine-tuning a small language model on my own emails and it started writing like I'm way too formal

I fed it about 200 of my work emails from the last 3 months and the output came back sounding like I'm a 1950s secretary, has anyone else run into their model overcorrecting on tone like that?

2 comments

2 Comments

mary_martin2215d ago

Did you try pruning out all the "per our conversation" stuff first? Might just be your own habits messing with the data.

victor_jones9915d ago

Actually 200 emails was about the same number I used when I tried this and I thought the formal tone was fine at first. Then I realized the model was just picking up on all the "per our conversation" and "please find attached" phrases I use without thinking. It really changed my mind about how much our own writing habits can skew the training data.