Multi-value tag field

A multi-value tag field stores a list inside one cell - the standard way open product databases keep several classifications of the same item without widening the table. Open Beauty Facts records its granularity as 'multi-value tag fields for categories, brands, ingredients and countries', so one product row carries many brands, a stack of category tags, an ingredient list and market codes at once. Values are namespaced: the Hugging Face beauty split shows countries_tags entries like en:france, where en: marks label language rather than country. Because delimiter counts vary row to row, none of these fields arrive analysis-ready - every consumer must explode the lists before counting, while the sku-barcode-record grain underneath stays one row per barcode.

Treat a packed tag column as a clean categorical field and every downstream count is wrong. A brand filter applied to a raw countries_tags-style cell matches only single-value rows, so coverage looks far thinner than it is. Aggregating across the en:-prefixed namespace without stripping it splits one country into several apparent markets. And because each record repeats the full ingredient list, joining on a substring silently duplicates rows - the failure mode that turns a product-count query into an inflated inventory figure. The cost is real work before first insight, not just inconvenience.

1. Confirm the field is genuinely multi-valued. Open Beauty Facts declares multi-value tag fields for categories, brands, ingredients and countries, so expect lists, not scalars, in all four.
2. Inspect real values before writing a splitter. The Hugging Face beauty split shows observed countries_tags values such as en:france, so your parser must handle language prefixes, not just delimiters.
3. Explode, then count. Unnest each list into one-row-per-value form so brand and country frequency counts are computed on values rather than cells.

Entries that sit next to multi-value tag field in this glossary:

• nightly-bulk-dump — A complete cumulative snapshot regenerated every night and published as JSONL, CSV, Parquet, RDF and MongoDB downloads.
• sku-barcode-record — One row per retail barcode, so each physical SKU is its own record.

The ingredient-function-tag entry - the formulation role assigned to each ingredient - sits in the same beauty-data slice, alongside household-products data and personal-care-products data.