Motivation is interesting. They saw that recent LLMs did much better in translating books than Google Translate does, but they did not want the inputs being used as training data, especially as they had documents they “prefer not to make publicly available online”.
The contribution here is an interface to translate a book that makes use of existing multi-language capable Llama model.
One comment of feedback is on prompting technique for translation—asking the model to translate results in lower quality than asking to rewrite the text as if it was written by a native writer.
One interesting thread of discussion is the OP suggesting that while they are the one who created this tool, there are nuances to translating books like translating invented terms (i.e., “facecrime” in 1984) into another language, which will involve the LLM also inventing that term in the output language. However, when this happens, discussion with the original author on whether the new word in the output language is indeed what they intended to mean in the original work. As a reply, another commenter says they do not think human intervention would be needed if there was a big enough LLM and good prompt engineering.