Hi,
how would you do a finetune job so it’s capable to answer in several languages afterwards?
Meaning if I do a finetune for lawyers, but let’s say in english and I do my question in spanish?
best
For me there is no issue in legal domain, fine-tuned in french, works in English and Spanish…
BUT, fine-tuning is not for question answering (from my experience) ESPECIALLY IN LEGAL DOMAIN.
I’m doing legal document analysis. For me answering a simple question about a fact from the document or whether this or that requirement is fulfilled by the document is a result of some 30 operations running on 2 databases and fine-tuned models, one of the databases has over 60 tables…
Answering precision is close to 100% (over 99,8), and still there is a huge disclaimer where I show that all analysis results are based purely on user custom configuration of the tool and are not a legal advice in any form but a mere highlight of potential issues to help the user take their own decision.
So please, make sure you understand where you’re going with this.
It’s just an example. We also do RAG.
But generally speaking,
sometimes the answer ends up being completly mixed up even some russian letters show up.
Might be worth checking encoding in the training files and mismatches between languages in system/user messages vs assistant’s replies.