This research will present a grounded and systematic improvement in classifying outcomes of cases by application of n BERT fine-tuning using the augmented (Legal) BERT model on a large-scale legal text of extensive imbalanced data. Also, through class imbalance control through oversampling, along with the calculation of the respective weights of classes, foreseeing the promotion of model generalization can also be said. A tailored pipeline for the dataset that adopts windowing tokenization with sliding windows is developed, with combined considerations to make start-up in tokenization, while at the same time, restricted by input length. For this study, the model presented already scored 94.5% training accuracy after six epochs and 54.2% validation accuracy which point toward an encouraging promise of significantly harnessing in-built transformer-based architectures for legal text classification. But there may be clues that the high level of generalization of these transformers acts toward unseen data. However, there is more necessity for accuracy and optimization analysis using differing transformer-based models. AI promises to automate an obvious extent of the legal text analysis with its tremendous capability and uncovers what the challenges would be for handling to a great extent imbalanced and complex datasets in the legal regime.