Trending datasets on Hugging Face (weekly)

Weekly popularity-velocity ranking of trending datasets on the Hugging Face Hub. Two sweeps of the official keyless Hub datasets API (https://huggingface.co/api/datasets), one sorted by likes and one by downloads (top 2 pages each), are merged and deduped on dataset id; gated, disabled and private datasets are dropped. Each dataset carries its license tag normalized to a canonical id (informational, not legal advice), its task_categories tags normalized to a coarse modality (text, vision, audio, video, multimodal, other, or unknown when no signal) via a documented priority rule set, its task categories, size category, formats and region tags extracted into dedicated columns, and author/name parts. Popularity velocity is likes and downloads per day since dataset creation (age floored at 1 day); the trend_score is a documented 0-100 composite = 50% min-max-normalized likes/day + 50% min-max-normalized downloads/day, and trend_rank orders by trend_score descending (ties: likes desc, then dataset id). Columns: ISO week, fetch timestamp, dataset id / URL / author / name, modality, task categories, size category, formats, region, normalized license id, comma-joined tags, likes, downloads, creation timestamp, age in days, likes/day, downloads/day, trend score (0-100), trend rank. Primary key: (week, dataset_id). Cadence: weekly. Nullability: task categories, size_category, formats, region and license_spdx may be empty when the author supplied none; likes, downloads, trend_score and trend_rank are never null. Caveats: likes/downloads are cumulative totals, so per-day rates are lifetime averages, not trailing-week gains; the rank therefore favors datasets with sustained momentum as well as fast risers; license_spdx comes from author-applied tags and may be missing or wrong — verify before use; popularity is not quality. Sample use: order by trend_rank for the week's hottest datasets, or filter modality = 'vision'.