Saidzoda-Tajik-LFM-2.6B Improves Tajik Tool Calling Accuracy

Adapting a 2.6B hybrid model via LoRA continued pretraining and SFT increases Tajik tool-calling accuracy from 20.69% to 64.14%. The study reveals that continued pretraining on specific languages can inadvertently degrade native tool-calling capabilities in English.