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

In this tutorial, we develop an end-to-end sentiment analysis workflow using the Stanford NLP IMDb Large Movie Review Dataset and compare classical machine learning with parameter-efficient transformer fine-tuning. We begin by establishing a reproducible environment and auditing the dataset for class ordering, review-length skew, duplicate leakage, and preprocessing artifacts before training a strong TF-IDF and Logistic Regression baseline. We then fine-tune DistilBERT with LoRA through PEFT, evaluate it using accuracy, macro-F1, ROC-AUC, confusion matrices, and ROC curves, and examine threshold selection and probability calibration through Expected Calibration Error and reliability analysis. Beyond headline metrics, we investigate confident errors, performance across review lengths, word-level occlusion saliency, and head-versus-tail truncation to understand how the model reaches its predictions and where long-context limitations affect performance. Finally, we use the unlabeled IMDb split for confidence-based pseudo-labeling, compare the resulting semi-supervised model against our baseline, and save the merged transformer for reusable sentiment inference.

We configure the Colab environment, install the required libraries, apply the PEFT–torchao compatibility fix, and set deterministic seeds for reproducible experiments. We load the Stanford IMDb dataset, shuffle and subsample the train and test splits, and inspect class balance, review-length distributions, duplicate leakage, and HTML artifacts. We also visualize review lengths and label frequencies so we understand the dataset structure before building any models.

We train a strong TF-IDF and Logistic Regression baseline and inspect the most influential positive and negative n-grams to establish an interpretable reference point. We then tokenize the IMDb reviews and configure DistilBERT with LoRA adapters that update only a small subset of model parameters while keeping the backbone largely frozen. We use the Hugging Face Trainer with dynamic padding, early stopping, mixed precision, and multiple evaluation metrics to fine-tune the transformer efficiently.

We evaluate the fine-tuned DistilBERT-LoRA model using classification metrics, a confusion matrix, and ROC curves while directly comparing its ROC-AUC performance with the TF-IDF baseline. We sweep classification thresholds to determine whether the default probability cutoff of 0.5 gives the best accuracy on our evaluation set. We also calculate Expected Calibration Error and construct a reliability diagram to measure how closely the model’s predicted confidence corresponds to its actual correctness.

We examine the model’s most confident incorrect predictions and group reviews by length to identify truncation-related failure patterns and difficult examples. We merge the LoRA adapters into the underlying model and apply leave-one-word-out occlusion to estimate which words push individual predictions toward positive or negative sentiment. We then compare predictions based on the beginning and ending portions of long reviews to determine where the strongest sentiment information resides.