The concept of "machine unlearning" has rapidly emerged as a critical trend in Machine Learning, driven by increasing regulatory demands for data privacy and the "right to be forgotten." Defined as the process of efficiently removing the influence of a specific training data instance from a trained machine learning model without necessitating a complete retraining from scratch, this capability is crucial for compliance and maintaining data governance. While "exact machine unlearning" focuses on techniques guaranteeing complete removal, current state-of-the-art approximate unlearning approaches often fall short, degrading representational quality or merely modifying the classifier without genuinely eliminating the targeted data from deeper representations. To address these limitations, new frameworks like Sequence-aware Sharded Sliced Training (S3T) are being developed, utilizing a lightweight parameter-efficient fine-tuning approach that enables parameter isolation. This research is vital for building truly responsible, auditable, and privacy-preserving AI systems in an era of increasing data sensitivity, requiring a fundamental alteration of the model's learned knowledge.