Daniel van Strien trains a 194M GLiNER2 model to suggest Hugging Face dataset task tags

Daniel van Strien trained a Jev-style classifier on Hugging Face Jobs for about $1.50. It is a 194M GLiNER2 model that suggests task tags for any Hugging Face Hub dataset from its column names and first row, and returns a label with a probability.

He reports that GLiNER2's first zero-shot suggestion matched an owner's tag 10% of the time, and 69% after 17 minutes of fine-tuning. The fine-tuned model runs on a free CPU in about a second. He says owners' tags are noisy, so some answers counted as wrong are tags the owner left out.

The recipe is open: one hf jobs command trains the same kind of model on a user's own labels. The README example, using book titles, runs in about two minutes for about $0.02. The demo is the hub-task-tagger Space, and the recipe is the uv-scripts classification dataset.