FixBerry

Fixberry is a little dataset I have made to train models to correctly count the number of letters in a word. It is commonly known that even the best LLMs fail at counting the number of R's in strawberry. I have also found out they have problems with other words too, like keeper and parallel but weirdly not with words like pepper and peeper. This should really be investigated more closely, I suspect it has something to do with tokenization and the possability that the model is only counting up to 1 letter per token, even if the token contains 2 letters.

That being said, I HIGHLY recomend not using this dataset, if you can fix the core problem, please do that instead!

The dataset was made from a list of english words from https://www.kaggle.com/datasets/lennartluik/all-english-words-csv I processed them into what you see in Data.csv