Hy-MT2-1.8B for WMT26: Staged Multilingual Adaptation for Chinese–Southeast Asian Translation

Abstract

We describe a training and evaluation pipeline for Hy-MT2-1.8B in the WMT26 Chinese– Southeast Asian multilingual machine translation task. The system uses one instruction-following model for fourteen directions be-tween Chinese and Indonesian, Thai, Khmer, Malay, Vietnamese, Burmese, and Lao. Our pipeline separates three responsibilities: mono-lingual continued pretraining (CPT), bilingual supervised fine-tuning (SFT), and evaluation-time serving. We construct leakage-controlled data with a complete chat-template token-length check, train a shared CPT checkpoint, and compare parameter-efficient LoRA SFT against full-parameter SFT under matched data, seed, and evaluation conditions. On the completed 26,474-example validation set, full-parameter SFT achieves higher macro-average sacreBLEU, COMET, and Translation Quality Score (the official WMT26 metric) than LoRA SFT, although the magnitude of improvement varies across directions.