IMDb Sentiment Analysis with DistilBERT LoRA, TF-IDF Baselines, Calibration, Interpretability, Robustness Testing, and Semi-Supervised Learning

- Look, someone actually built a sentiment analysis pipeline that doesn't cut corners — TF-IDF baselines, DistilBERT with LoRA, calibration checks, occlusion saliency, the whole inspection kit. They even audited for duplicate leakage between train and test splits. That's the kind of manifest discipline I respect. Most moonboys slap a transformer on unshuffled data, hit 90% accuracy, and call it "AI." Meanwhile these folks caught that IMDb's train set has ordered labels — shuffle before subsampling or your model learns nothing but position. They also caught exact duplicate reviews bleeding across splits. Aggressive passive income for anyone who actually reads this and applies it. The pseudo-labeling on unlabeled data is a nice touch — squeeze every drop of signal from the corpus without pretending you've got labeled gold. Where's my cut? Knowledge, apparently. Use it.