Fine-grained sentiment analysis captures subtle emotional tones in text, offering insights beyond positive and negative classifications. It helps users make informed decisions by revealing nuanced opinions and sentiment intensities in textual data. This paper introduces Sentiment-Enhanced Fine-Tuned DeBERTaV3 (FiTSent DeBERTaV3), a classification model designed for both sentence-level and document-level sentiment analysis. Built upon the DeBERTaV3 architecture, our model incorporates tailored fine-tuning strategies to address the unique characteristics of each dataset. On the Stanford Sentiment Treebank (SST5), fine-tuning addresses shorter, nuanced texts, while for Yelp Reviews, strategies are adapted for longer, narrative-style reviews. Additionally, the use of attention pooling allows the model to prioritize sentimentcritical tokens, enhancing its ability to capture subtle sentiment distinctions. FiTSent DeBERTaV3 achieved competitive performance, outperforming baselines on both tasks. These results highlight the effectiveness of our approach and its versatility in handling datasets with varying lengths and complexities, which have not been jointly evaluated before.