Lightweight Baselines for Medical Abstract Classification: DistilBERT with Cross-Entropy as a Strong Default

Large language models work well for many NLP tasks, but they are hard to deploy in health settings with strict cost, latency, and privacy limits. We revisit a lightweight recipe for medical abstract classification and ask how far compact encoders can go under a controlled budget. Using the public medical abstracts corpus, we finetune BERT base and DistilBERT with three objectives standard cross-entropy, class weighted cross entropy, and focal loss keeping tokenizer, sequence length, optimizer, and schedule fixed. DistilBERT with plain cross-entropy gives the best balance on the test set while using far fewer parameters than BERT base. We report accuracy, Macro F1, and Weighted F1, release the evaluation code, and include confusion analyses to make error patterns clear. Our results suggest a practical default: start with a compact encoder and cross-entropy, then add calibration and task-specific checks before moving to heavier models.