Topic Feed Lab

A small open model, fastino/GLiNER2.5-Decide, reads each item once and gives it a score from 0 to 1 for each of 48 topics, in one call. It runs on CPU. It is a classifier, not an LLM: it does not write text.

• What is scored: new and updated public Hub models and datasets that have a real card (at least a few sentences), and papers from Hugging Face Daily Papers. The first week also includes arXiv computer-science papers from a one-off backfill. Quantized and converted models take their base model's topics. Items with no README or an empty template card are skipped.
• When it shows up: a worker process in this Space follows the Hub, checks whether each card changed, scores changed items and stores the scores. Models usually appear within a minute or two of the change; datasets are found by a slower sweep and can take up to about half an hour. The page reads the stored scores.
• Thresholds: an item is tagged with a topic when its score reaches that topic's threshold. The defaults are set so each topic tags about as many items as a larger LLM did on a reference sample.
• Your confirmations: tick or untick the topics on an item and press Confirm. After 3 ✓ for a topic, its threshold is fitted to your confirmations. You can also set it by hand under calibrate. This is stored in your browser only.
• Caveats: the topics are zero-shot, so some tags are wrong, and broad topics catch loosely related items. The model also gives “historical newspapers” high scores for general news, so that topic counts only when the item's name, card summary or tags have a historical marker such as “archive” or a year from 1500 to 1999. Likes and downloads are a snapshot from when the item was scored.