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.
