Sentiment analysis in low resource multilingual contexts, such as Javanese in Indonesia, presents challenges due to ambiguous grammar and vocabulary. In addition, social media users in Indonesia frequently use mixed languages, commonly referred to as code mixed text, specifically combining Indonesian and Javanese. This study evaluates the performance and efficiency of Transformer based approaches for ternary sentiment classification by comparing the monolingual IndoBERTweet and multilingual XLM RoBERTa models. It also examines the effect of translating code mixed text into Indonesian during preprocessing and compares full fine tuning with Parameter Efficient Fine Tuning (PEFT) using the IA3 method. Experimental results indicate that translating text to Indonesian consistently improves both models, particularly for the neutral class, while clarifying confusion matrix prediction patterns. The configuration using IndoBERTweet, Indonesian text, and full fine tuning achieved the highest Macro F1 score of 0.8180, whereas applying PEFT to this setup yielded a score of 0.8027. For XLM RoBERTa with Indonesian text, PEFT matched and slightly exceeded full fine tuning with a peak Macro F1 score of 0.7986. Computationally, PEFT provides significant savings by reducing trained parameters by over 70 percent, cutting training time by around 57 percent, and shrinking the model checkpoint size by over 98 percent without substantial performance loss. Ultimately, combining Indonesian text preprocessing with PEFT proves to be an effective, efficient, and practical approach for developing sentiment analysis systems for multilingual digital communication in Indonesia.