Fine-Tuning an SLM for a Low-Resource Language

Oh. I don’t know Persian myself, but Gemma is basically like Gemini’s younger sibling, so my guess is that it may be fairly strong for Persian too. Also, leaderboards can make the search much easier: * Open Persian LLM Leaderboard / OPLL Open Persian LLM Leaderboard - a Hugging Face Space by opll-org * Another Open Persian LLM Leaderboard mirror / variant Open Persian LLM Leaderboard - a Hugging Face Space by PartAI * MIZAN: Persian LLM Leaderboard MIZAN: A Persian LLM Leaderboard - a Hugging Face Space by MCINext * ParsBench GitHub - ParsBench/ParsBench: ParsBench provides toolkits for benchmarking LLMs based on the Persian language tasks. · GitHub ParsBench (ParsBench) * FaMTEB, if embeddings / RAG are relevant [2502.11571] FaMTEB: Massive Text Embedding Benchmark in Persian Language * PersianPhi may also be worth checking as a Persian-adapted compact model amirakhlaghiqqq/PersianPhi · Hugging Face So, if me, I’d probably use those leaderboards to make a shortlist, then compare tokenization and a small private Persian eval before choosing the base model.