AIDataset tagger reportedly trained for about $1.50 on Hugging Face Jobs
A user says 17 minutes of fine-tuning raised GLiNER2's first-suggestion match rate with an owner's dataset tag from 10% to 69%. They caution that owners' tags are noisy, so some apparent errors reflect tags the owner left out.
TLDR
A user shared a demo and open recipe for a GLiNER2 classifier that suggests task tags from a Hugging Face Hub dataset's column names and first row, returning a label with a probability. They report training it for about $1.50: its first suggestion matched an owner's tag 10% of the time without fine-tuning and 69% after 17 minutes of fine-tuning. The user cautions that owners' tags are noisy and sometimes incomplete. They say the tuned model runs on a free CPU in about a second, and the recipe uses one hf jobs command to train the same kind of model on custom labels.
