Ichlibitiche/incidb-skincare-free-sample

INCIDB: Skincare & Cosmetics INCI Database (free sample)

Full dataset: incidb.dataengineered.io · $79 one-time (INCIDB Complete, CSV + Parquet) → Buy on Stripe · the same sample on Kaggle

This is the free sample, not the full corpus: 200 products drawn at random from a seeded eligible pool, with their brands, the 1109 ingredients they reference, all their composition links, and the matching name-map rows. Identical schema and identical columns to the paid snapshot.

A normalised relational snapshot of what is actually printed on cosmetic ingredient labels. The full snapshot (18,583 products, 5,925 brands, 46,973 distinct canonical INCI names, 318,758 ordered composition links, 55,426 name-map rows) is at the portal below.

Portal, schema and measured coverage: https://incidb.dataengineered.io/

---

Tables

products, brands, ingredients, product_ingredients and ingredient_name_map. Composition links preserve the label order in position_index, so relative concentration rank is recoverable.

Sources

Two, joined explicitly:

1. Open Beauty Facts — the product corpus (barcodes, brands, the source's own categories_tags, raw on-pack ingredient declarations), under the Open Database License (ODbL).
2. European Commission CosIng — the ingredient inventory: functional categories, CAS and EC numbers, chemical descriptions, restrictions, Annex II–VI membership. Joined on exact canonical name; unmatched names are NULL in every CosIng-derived column, never a placeholder.

Allergen flags come from one list only: the EU Annex III fragrance allergens (Regulation (EC) 1223/2009 as amended by Regulation (EU) 2023/1545) — 99 names flagged corpus-wide. The US FDA has not yet published its MoCRA fragrance-allergen list, so this dataset carries no US flag. comedogenic_rating covers 145 ingredients transcribed from Fulton JE Jr., J Soc Cosmet Chem 1989;40:321–333 (Table I) and is NULL elsewhere. is_fungal_acne_trigger is a rule-derived heuristic (270 flags corpus-wide), not a measured property.

Measured coverage — read this before modelling

Corpus-wide, stated two ways because they differ by an order of magnitude:

Column	Share of the 46,973 distinct names	Share of the 318,758 label occurrences
CosIng match (any)	11.0%	82.5%
functions	10.9%	81.5%
cas_number	8.3%	76.4%

A label corpus contains far more distinct strings — botanical variants, multilingual spellings, marketing tokens — than any regulatory inventory lists, which is why the left column is low; the slots on a real label are filled overwhelmingly by ingredients CosIng does cover, which is why the right column is high.

The sample reads better than the corpus average by construction: products were only eligible for sampling when at least 80% of their linked ingredients are CosIng-matched. Plan against the corpus figures above, not against this sample.

ingredient_name_map — the canonicalisation audit trail

Every raw label token, the canonical INCI name it resolved to, and the method that resolved it (exact, cleaned, typo_map, percent_stripped, paren_stripped, synonym_map, slash_variant, split, or unresolved). Nothing that failed every rule was guessed at — it is kept verbatim and marked unresolved.

Ideas

• Ingredient scanners: resolve a scanned label through the name map, then surface CosIng functions and Annex III flags for the slots you can resolve.
• Composition models: position-weighted ingredient vectors across the products.
• Regulatory analysis: Annex II–VI membership and restrictions by brand.

Attribution & licensing

Product data © Open Beauty Facts contributors, licensed under the Open Database License (ODbL) v1.0 — derived databases are subject to attribution and share-alike. Ingredient enrichment contains data from the European Commission CosIng database, reused with attribution. Provided as-is; the flags and ratings are informational, not medical or safety advice. Questions: incidb@dataengineered.io.

Quick start

All files are pipe-delimited (sep="|") UTF-8 CSV.

import pandas as pd

base = "hf://datasets/Ichlibitiche/incidb-skincare-free-sample/"
products = pd.read_csv(base + "products.csv", sep="|")
ingredients = pd.read_csv(base + "ingredients.csv", sep="|")
links = pd.read_csv(base + "product_ingredients.csv", sep="|")
name_map = pd.read_csv(base + "ingredient_name_map.csv", sep="|")

Links

• Portal, schema, measured coverage and the paid snapshot ($79 one-time): https://incidb.dataengineered.io/
• Kaggle mirror of the free sample: https://www.kaggle.com/datasets/dataengineered/incidb-skincare-and-cosmetics-inci-formulations
• Live demo Space: https://huggingface.co/spaces/Ichlibitiche/incidb-skincare-dupe-finder
• Questions: incidb@dataengineered.io