Model choice for Persian projects is usually made from English leaderboards — but Persian errors are different in kind: zero-width non-joiners, numerals, administrative terms and long sentences.
Build your own test set
Collect one to two hundred real questions from your domain, write the correct answer for each, and add some deliberately hard ones — Persian dates or rial amounts, for example.
Four criteria
Answer correctness, faithfulness to the source (in RAG), Persian language quality and instruction-following. Score each separately so you know which weakness fine-tuning must fix.
After fine-tuning
Run the same test set again. A before-and-after report is one of the deliverables of our enterprise LLM projects, because without it nobody can judge whether fine-tuning was worth it.