Code-mixing refers to the blending of languages. The most serious issue with code-mixing is that people move between languages and type in English instead of transcribing words in Dravidian scripts. Traditional NLP models are trained on large monolingual datasets. Code-mixing is difficult since they can’t manage data code-mixing. A significantly less percentage of research is going on code-mixed data. In this paper, a Kannada-English code-mixed dataset is used. In this dataset, YouTube comments are categorized into a negative, positive, unknown state, not-intended language (not-Kannada) and mixed feelings, and categories based on message-level polarity. The summation of hidden layers of the pre-trained model (SHLPM) like BERT and XLM-RoBERTa classification models was used for this complex and challenging task. The SHLPM-XLM-RoBERTa has given F1-score of 64% for the Kannada-English YouTube comments dataset and outperforms other existing models.