adaptive-classifier

Why we included this project

Text classifiers in production rarely stay static: the label set grows, new examples show up, and full retraining is disruptive. Adaptive Classifier is a PyTorch library that tackles that directly, letting you add classes and training examples at runtime without catastrophic forgetting through elastic weight consolidation and a FAISS-backed prototype memory, on top of any HuggingFace transformer. Model updates happen without downtime, and the built-in ONNX Runtime path speeds up CPU inference, so it fits live services rather than just experiments. There's also a game-theoretic prediction mode that resists adversarial or strategically manipulated inputs, which matters for moderation and support routing where users try to game the system. If your data keeps moving after deploy, this is a focused way to keep the classifier in step with it.