Explosion just shipped Prodigy-HF, a plugin that closes the loop between annotation and model training. If you’ve been labeling data in Prodigy for named entity recognition or text classification, you can now fine-tune any transformer model from the Hugging Face Hub against that data directly from the command line. No export scripts, no glue code, no wrestling with tokenizers.
The mechanics are refreshingly simple. Install the plugin, run a command like python -m prodigy hf.train.ner fashion-train,eval:fashion-eval path/to/model-out --model "distilbert-base-uncased", and Prodigy handles the rest. The plugin taps into Hugging Face’s AutoTokenizer and AutoModel classes, which means the entire model hub is fair game. That flexibility matters: a team working in a niche domain can grab a multilingual model or a domain-specific checkpoint and fine-tune it on a few hundred of their own annotations without leaving Prodigy.
The second half of the plugin tackles the collaborative side of annotation. A new hf.upload command pushes annotated datasets straight to the Hugging Face Hub, letting teams share labeled data with the wider community. Explosion credits Hugging Face’s @davanstrien for suggesting the upload feature. There’s also an hf.correct.ner recipe that lets you reuse a fine-tuned NER model inside Prodigy’s annotation interface, with model predictions displayed alongside your labels — a meaningful time-saver for iterative labeling workflows.
What makes this worth paying attention to is how it lowers the barrier between data work and model work. The two have traditionally lived in separate tools with brittle handoffs. Prodigy-HF makes the round trip — annotate, fine-tune, re-annotate with model assistance, share the dataset — feel like one continuous workflow rather than a pipeline you have to babysit. Explosion says more plugin features are on the way, and given how much this one was requested on their support forum, the demand signal is clear.
