qwen3-1.7b-sunda

LoRA fine-tune of Qwen/Qwen3-1.7B for Indonesian / Sundanese / English translation.

metric	score
chrF	52.7
BLEU	27.1

What these numbers mean

Trained on NusaX-MT (3800 train / 200 test pairs), built by pairing the Sundanese column against Indonesian and English in both directions. SEACrowd/parallel_su_id could not be loaded (datasets 4.x dropped script-based datasets), so NusaX is the entire training set.

About 4,000 pairs is very little for translation. Treat this as a demonstration that a small Qwen3 can be taught usable Sundanese, not as a production system.