Hi.
Is there any consensus on how one can effectively use LLMs for professional translations, like some documents?
Is it even possible to have an LLM that is capable of translating single documents from language A to language B with no problems, I.e having outputs that are 100% correctly translated docs?
There are various benchmarks for measuring the performance of LLM and embedding models, but translation tasks are difficult to benchmark. This is especially true as the task becomes more refined. While embedding models are relatively easy to evaluate, the level required for LLM is high. This is because even if the meaning is correct, the translation must also be natural to achieve 100% accuracy. However, it is possible to estimate the basic performance of each model architecture to some extent using benchmarks.
Regarding the individual translation capabilities of models, models fine-tuned for the target language generally become capable of producing translations that feel natural in that language. Training specifically tailored to translation tasks could yield even more effective results.
For major languages other than English, there may be language-specific LLM leaderboards available, which could serve as a reference. Since all models typically have English understanding capabilities as a default, they should be suitable for translation tasks as well.