One commit per update.
Each commit to a Hugging Face dataset restarts its viewer, the table preview on the dataset page, for about 33 minutes. One data update made two commits, tables then search vectors, and a card edit made a third. The publish step now waits for the same commit's vectors and uploads tables, card, methods, schema and vectors together.
A card you can start from.
The card now has the same opening as the GitHub README: what the library is, live counts and a screenshot. Then come four ways to start in one minute: the datasets library, a DuckDB query that reads the Parquet files straight from Hugging Face, the MCP server and the API. A table lists the 10 main tables and links the full schema.
Badges that showed on GitHub and nowhere else.
The live count badges were blank on Hugging Face. Our site told browsers its files were for our site only, and Chromium on huggingface.co blocked all five. GitHub had hidden the problem because it fetches images through its own proxy. Badges now say any site may show them. Then they stacked one per line, because the Hub styles a linked image as a block, and one inline style put them back in a row.
An older run overwrote a newer card.
The badge fix published at 13:28 UTC. A data run from a commit merged just before it finished at 13:44 and uploaded the card as that older commit had it. The data publish now takes the card from the main branch at the moment it uploads, so a slow run can no longer bring back an old card.