In Bangladesh’s rapidly evolving educational landscape, social media platforms like Facebook have become pivotal for informal teacher–student engagement, especially in university admission coaching. This study analyzes teacher feedback from Facebook comment sections under Bangla-language educational videos, offering insights into how digital feedback shapes learning dynamics. A new dataset, BnAdEduFeed3, was developed by scraping and manually annotating 6,010 comments across three sentiment classes: Positive, Neutral, and Negative. After preprocessing, Word2Vec-based synonym replacement was applied for augmentation to improve lexical diversity and class balance. Three transformer models: sagorsarker/bangla-bert-base, csebuetnlp/banglabert, and xlm-roberta-base, were fine-tuned, with bangla-bert-base achieving the best results (accuracy: 95.97%, precision: 93.97%, recall: 96.97%). The findings establish a benchmark for sentiment analysis in Bangla educational discourse and highlight the pedagogical potential of real-time, student-generated feedback. Ethical safeguards, including anonymization and adherence to platform guidelines, were maintained throughout. While the dataset’s specificity limits generalizability, the study advocates extending research to multi-platform and multi-modal feedback in low-resource language contexts, reinforcing Facebook’s role as a feedback ecosystem in modern education.