Language Models for Text Classification: From Bag-of-Words to Jev
Summary
The article surveys the landscape of text classification, tracing the evolution from bag-of-words to Jev. It covers classic methods (Naive Bayes, logistic regression), word embeddings, RNNs and CNNs, transformer architectures (BERT, GPT, T5), and introduces Jev with its Choice, Noul, and Score APIs. It compares Jev to GPT-style models in terms of latency and cost, discusses calibration and RLCD-based training, and reviews experimental results on IMDb alongside broader use cases and ecosystem activity. The piece also touches on Jev clones and industry developments, including OpenAI’s Decision API.