Tesseract training: just a few words

I need to train tesseract to recognize only ten words.
The words are the name of the farm, for example: Atrasil, Spectful
Since the fonts used are quite common, I tried to unpack eng.traineddata, replacing freq-dawg and word-dawg with these words. Then I repackage them into new trainable information, unfortunately it doesn't work very well. The matching results are still unacceptable and I cannot use them even when I use images obtained from a simple word file.
Is there a way to achieve a good match? do I need to remove other files from the unpacked data?