Social media is now recognized as the key medium for expressing thoughts on topics like interests, hobbies, and reviews, often using natural language involving code-mixing and script-mixing, where multiple languages and scripts are combined in informal, casual, and sometimes non-standard forms. This work addresses substantial proportion of users who communicate in regional languages, specifically Marathi. This paper presents MCodeScript, suite of unsupervised and supervised datasets for Marathi code-mixed and script-mixed data. MCodeScript-un is an unsupervised dataset of ~1.3 L comments (~1.5 M tokens) gathered from diverse sources, including social media, community organizations, and news websites. We also present MCodeScript-MeSent, which transforms the MeSent supervised Romanized Marathi dataset to the script-mixed form using our automated technique. We fine-tune MCodeScript-MeSent for sentiment analysis tasks using two pre-trained BERT-based models, mBERT and MeBERT-Mixed-v2. The fine-tuned models, MCodeScript-mBERT and MCodeScript-MeBERT, achieve F1 scores of 53% and 67%, respectively, with MeBERT demonstrating a 14% improvement in prediction accuracy. Additionally, complexity and mixing levels of the MCodeScript-MeSent dataset are analyzed using code-mixing index (CMI). Also, this paper introduces Script-Mixing Index (SMI) to assess the extent of script inclusion in manually transformed datasets, accompanied by qualitative evaluation using readability and coherence metrics. Despite challenges like the informal nature of user-generated content, this paper contributes to the growing demand for linguistic evaluation of code-mixed content in both Roman and native scripts. The MCodeScript datasets facilitate the development of more effective language processing models, marking a significant step towards enhancing the accuracy as well as effectiveness of NLP approaches with Marathi-English code-mixed texts. All datasets are publicly released at https://data.mendeley.com/datasets/zcgbnn8zbm/1 .