This project fine-tunes BERT (bert-base-uncased) for a 3-class sentiment classification task using the syedkhalid0/Sentiment-Analysis dataset from HuggingFace, which contains ~105K English text samples labeled as Negative (0), Neutral (1), and Positive (2), split across train (84K), validation (10.5K), and test (10.5K) sets. The dataset covers diverse text sources including tweets, product reviews, and movie reviews, pre-processed to remove duplicates and null values for consistency. Built using HuggingFace transformers and datasets libraries, the pipeline handles tokenization via AutoTokenizer, dynamic padding via DataCollatorWithPadding, and training via the Trainer API — making it easy to reproduce or adapt for similar NLP classification tasks.
