predicts influenza a host species — human, avian, or other mammal — from the pb2 protein sequence using the esm-2 language model. reached 90.97% test accuracy (chance = 33%) with a random forest over frozen embeddings, and reports honestly that fine-tuning the transformer did not beat it, analyzing why rather than hiding the negative result.
an interactive scatter "compass" for 300 columbia culpa course reviews. zero-shot classification (bart-mnli) places each review on a professor ↔ course axis while a roberta sentiment model sets the positive ↔ negative axis, shipped as a published plotly graphic for spectator.