Fine Tuning for Style Transfer on Minimal Data Set

I want to fine tune on a small data set of just 10 full length essays ranging from 500 to 1200 words. The hope is to get a model that generates completions in my style, tone, and voice while also being undetectable by common ai detectors. Having a hard time getting results that dont hallucinate, was using logit bias in addition to the fine tune but got crazy hallucination. I get on average a 8% vocab overlap and around 8% cosine similarity between training data set and the completions. Ive recently tried using NLTK to split my data set into smaller context aware snippets to train on more data, but results have not been as good. Ideally the smaller the dataset the better, and Im not using prompt completion pairs just completions with a system message in the fine tune. Any tips or ideas for making this better and consistent?

1 Like