Hugging Face’s AutoNLP framework just proved it can handle multi-class text classification with almost absurd ease. The team took the BBC News Classification dataset from Kaggle—roughly five categories: business, entertainment, politics, sport, and tech—and fed it into AutoNLP. No code. No hyperparameter fiddling. Just upload the CSV, select auto-splitting, and let the platform train 15 different transformer models simultaneously.
The result? A best model scoring 98.67% accuracy after about 15 minutes of training. That’s not a typo. For context, hitting near-99% on a five-class news categorization task would’ve required serious feature engineering and model selection effort just a few years ago. AutoNLP handles it as a background process while you go get coffee. Pricing starts around $10 per model, which makes this accessible for teams that don’t have the budget for dedicated ML engineers.
But the more interesting part of this project is what happened next. The dataset had categories, sure, but it lacked named entity labels. So the team used Prodigy—Explosion’s annotation tool built by the spaCy creators—to label entities like PERSON, ORG, PRODUCT, and LOCATION. Prodigy’s web interface runs locally, typically at localhost:8080, and serves up unlabeled documents one at a time for manual tagging.
Here’s where the active learning angle comes in. After labeling just 20 samples, they trained an AutoNLP model on that tiny entity recognition dataset. This is the kind of workflow that makes NLP practitioners sit up: a commercial annotation tool feeding a no-code training pipeline. Prodigy doesn’t export in AutoNLP’s native format, so they wrote a conversion script using Prodigy’s database connector and spaCy. The implication is clear—active learning with small, strategically labeled datasets is becoming viable for teams without deep ML expertise. Whether 20 samples produces a useful NER model is an open question, but the pipeline itself is remarkably frictionless.
