PatMT2 Two-Stage LoRA for Multilingual Patent Translation
October 14, 2026
PatMT2 is a two-stage LoRA pipeline that adapts the Hy-MT2-7B model for Chinese, English, Japanese, and Korean patent translation. It achieves a macro-average BLEU of up to 43.57, outperforming Qwen3-32B in specific translation directions.
HOW THIS AFFECTS YOU
●
builderYou can use this specialized pipeline to improve translation quality in highly technical domains.
●
researcherThe study provides diagnostic insights into model collapse in low-resource languages during adaptation.