Neural Machine Translation (NMT) traditionally struggles with low-resource language pairs like Chinese–Vietnamese, especially when faced with informal teencode and slang. We tackle this challenge with a new pipeline that combines multitask fine-tuning with Parameter-Efficient Fine-Tuning (PEFT). We built a specialized corpus by expanding 50,000 seed sentence pairs into over 280,000 multitask training instances covering four distinct translation tasks. Using Low-Rank Adaptation (LoRA), we fine-tuned and conducted a detailed comparison of two major multilingual models: mBART-50 and NLLB-200. The models were evaluated using a comprehensive suite of metrics, including BLEU, COMET, chrF++, and TER. The results show a substantial leap in quality; mBART-50 proved more stable and adaptive to Vietnamese teencode, achieving 29.36 BLEU and 82.64 COMET on informal Vietnamese-to-Chinese translation, outperforming NLLB-200 (26.36 BLEU and 80.66 COMET). Notably, both models retained strong performance on formal tasks (over 41 BLEU), with NLLB-200 showing superior semantic preservation in Chinese-to-Vietnamese translation (42.37 BLEU). Our work confirms that this targeted, multitask approach is highly effective for handling non-standard language in low-resource translation.