FragDBnet/fragrance-database

FragDB v5.18 — Fragrance Database (Multilingual Sample)

The most comprehensive structured fragrance database available. This is a free sample of FragDB: 141,541 perfumes, 23 languages — 10-row CSV samples at root.

Full dataset: fragdb.net.

What's New in v5.18

• Data updated from v5.17 → v5.18 (snapshot 2026-10-08):
  • Fragrances: 140,946 → 141,541 (+595: 601 added, 6 removed because their pages were deleted at the source — PIDs 122906, 123027, 134700, 138184, 138572, 142797)
  • Brands: 8,359 → 8,384 (+25)
  • Perfumers: 3,144 → 3,154 (+10)
  • Notes: 2,610 → 2,615 rows in notes.csv (+5)
• Companion parquet refreshed — user reviews 4,988,692 → 5,022,398 (+33,706, all English), editorial articles 25,709 → 25,780 (+71), news comments 277,446 → 278,905 (+1,459).
• Brand ids now survive renames and merges at the source. A renamed brand keeps its id and takes the new name (b7118 Saleel → Saleel Parfums, b4760 Aromatic Families → Aromatic). When the source folds one brand page into another, both ids stay and now carry the same name and page URL — five pairs this release, the largest b532 (417 perfumes) with b8330 (15), Novaya Zarya – The New Dawn. Joins on brand_id are unaffected; group by id, not by name, to keep each pair apart. Full list in the CHANGELOG.
• Two perfumes moved to another brand because the source moved them: PID 115180 (Maison Incens → Rocco d'Orazio) and PID 141384 (Opula Perfumes → Opal Perfumes).
• The 84 articles that v5.17 shipped without a publication date are now dated; in v5.18, 23 of the 29 newest articles (NID 26,262 and above) carry date_unix = 0.
• Sample files unchanged: the schema did not move between v5.17 and v5.18.

What's New in v5.17

• Data updated from v5.16 → v5.17 (snapshot 2026-09-30):
  • Fragrances: 140,230 → 140,946 (+716)
  • Brands: 8,316 → 8,359 (+43)
  • Perfumers: 3,126 → 3,144 (+18)
  • Notes: 2,606 → 2,610 rows in notes.csv (+4)
• Companion parquet refreshed — user reviews 4,986,774 → 4,988,692 (+1,918), editorial articles 25,619 → 25,709 (+90), news comments 276,483 → 277,446 (+963).
• Figures on this page and in SPEC.md are measured against the v5.17 files. Several had been left behind by earlier releases — the snapshot dates, the category split, the PID references, the archived share and the reply rate here, and most of SPEC.md, which still described the May 2026 snapshot. All of them are now restated.
• Correction: 84 of the 250 newest articles (NID 25,970 and above) carry date_unix = 0; earlier text said every non-archived article is dated.
• Sample files unchanged: the schema did not move between v5.16 and v5.17.

What's New in v5.16

• Data updated from v5.15 → v5.16 (snapshot 2026-09-19):
  • Fragrances: 139,501 → 140,230 (+729)
  • Brands: 8,272 → 8,316 (+44)
  • Perfumers: 3,116 → 3,126 (+10)
  • Notes: 2,596 → 2,606 rows in notes.csv (+10)
• Companion parquet refreshed for the first time since June — user reviews 4,643,851 → 4,986,774 (+342,923), editorial articles 24,440 → 25,619 (+1,179), news comments 263,798 → 276,483 (+12,685).
• New: a Data API — per-record HTTP access to the same catalogue, priced per record. See the section below.
• Derived figures re-measured against the new files: review coverage 66.9% → 69.8%, English share 1.69M → 1.88M, reply rate 4.9% → 5.5%.
• Sample files unchanged: the schema did not move between v5.15 and v5.16.

What's New in v5.15

• Data updated from v5.14 → v5.15 (full source recrawl, parser run 260910):
  • Fragrances: 137,789 → 139,501 (+1,712)
  • Brands: 8,247 → 8,272 (+25)
  • Perfumers: 3,110 → 3,116 (+6)
  • Notes: 2,592 → 2,596 rows in notes.csv (+7 new, 3 retired)
• Three note IDs retired — the source merged case duplicates: n473 → n2661 (Heather), n653 → n2646 (Icing Pink), n813 → n2660 (Hazelnut Cocoa Spread). No fragrance references the retired IDs.
• Figures on this card restated from the release files. The notes figure is now the row count of notes.csv, the number fragdb.net shows; the full-database totals had stayed at v5.10. Review coverage and foreign-key figures are measured against this catalogue.
• Reviews, news and news comments — the same parquet files as v5.14.
• Sample files unchanged — the schema is identical to v5.14, so the 10-row CSVs were not rebuilt. A sample shows structure, not freshness.

What's New in v5.14

• Data updated from v5.13 → v5.14 (incremental delta, parser run 260901):
  • Fragrances: 137,147 → 137,789 (+642)
  • Brands: 8,210 → 8,247 (+37)
  • Perfumers: 3,102 → 3,110 (+8)
  • Unique note names: 2,586 (unchanged — the notes reference only moves on a full crawl)
• Sample files unchanged — the schema is identical to v5.13, so the 10-row CSVs were not rebuilt. A sample shows structure, not freshness.

What's New in v5.13

• Data updated from v5.12 → v5.13 (incremental delta, parser run 260819):
  • Fragrances: 136,682 → 137,147 (+465)
  • Brands: 8,175 → 8,210 (+35)
  • Perfumers: 3,090 → 3,102 (+12)
  • Notes: 2,586 → 2,588 (+2) — including Kiwano, which arrived with all 22 translations in its first cycle
• Clean delta: no perfume changed its canonical URL and none disappeared; field coverage flat across all 30 columns (max movement 0.22 pp)
• Free sample refreshed — 10-record CSV samples rebuilt from v5.13 data

What's New in v5.12

• Data updated from v5.10 → v5.12 (full source recrawl, parser run 260809):
  • Fragrances: 135,308 → 136,682 (+1,374)
  • Brands: 8,093 → 8,175 (+82)
  • Perfumers: 3,057 → 3,090 (+33)
  • Notes: 2,573 → 2,586 (+13) — 100% multilingual, verified per language
• 784 records restored — a resume-scan defect had frozen them since May; their votes, ratings and note pyramids are current again
• Free sample refreshed — 10-record CSV samples rebuilt from v5.12 data

What's New in v5.9

• Data updated from v5.8 → v5.9 (parser run 260710):
  • Fragrances: 134,022 → 134,577 (+555)
  • Brands: 8,000 → 8,036 (+36) — 154 brand names canonicalized (restored Fragrance(s) suffix; IDs stable)
  • Perfumers: 3,035 → 3,046 (+11)
  • Notes: 2,562 → 2,567 (+5)
  • URL hygiene: pyramid anchors single-domain, photo cache-busters stripped
• Free sample refreshed — 10-record CSV samples rebuilt from v5.9 data

What's New in v5.8

• Data updated from v5.7 → v5.8 (parser run 260701):
  • Fragrances: 133,392 → 134,022 (+630)
  • Brands: 7,953 → 8,000 (+47)
  • Perfumers: 3,020 → 3,035 (+15)
  • Notes: 2,559 → 2,562 (+3)

What's New in v5.7

• Data updated from v5.6 → v5.7 (parser run 260619):
  • Fragrances: 132,858 → 133,392 (+534)
  • Brands: 7,927 → 7,953 (+26)
  • Perfumers: 3,005 → 3,020 (+15)
  • Notes: 2,550 → 2,559 (+9)

What's New in v5.6

• Data updated from v5.5 → v5.6 (parser run 260609):
  • Fragrances: 132,124 → 132,858 (+734)
  • Brands: 7,881 → 7,927 (+46)
  • Perfumers: 2,988 → 3,005 (+17)
  • Notes: 2,533 → 2,550 (+17)
• Notes multilingual 100% coverage (was 99.8%) — 6 previously gap-filled notes now complete across all 22 languages
• Photo URL stability — source cache-buster query params stripped, eliminating phantom diffs across releases
• Schema unchanged from v5.5 — existing loaders work without modification
• Free sample files unchanged (10-record structure preserved)

What's New in v5.5

• Data updated from v5.4 → v5.5 (parser run 260601):
  • Fragrances: 130,949 → 132,124 (+1,175)
  • Brands: 7,815 → 7,881 (+66)
  • Perfumers: 2,968 → 2,988 (+20)
  • Notes: 2,522 → 2,533 (+11)

Schema unchanged from v5.4

All F column counts identical (30/54/42/55/27/25) — existing scripts work without modification.

From v5.4 (unchanged in v5.5)

• 23 languages — all labels, note names, accords, countries, statuses translated
• 9 non-Latin scripts for perfumer name transliteration
• translations.csv — vocabulary file (34 entries) for gender and voting labels
• Compact notes pyramid — note_id,opacity,weight (name/icon via notes.csv JOIN)
• Each note name variant has its own ID with translations
• Gender & voting fields use translation IDs for multilingual support

Snapshot freshness

• Data refreshed: 2026-10-08 (v5.18)
• Reviews, news, news comments (parquet): refreshed in this release — latest review 2026-10-08, latest article 2026-10-06, latest news comment 2026-10-08

Data API — per-record access

The whole catalogue, one record at a time. A metered HTTP API over the same database we sell as files. You pay per record returned, at the level of detail you ask for. No subscription and no seats: top up a balance, fetch what your application needs, cache it as long as you like.

fragdb.net/api → · OpenAPI spec · Sign up · FAQ · Terms

Endpoints

GET  /api/v1/fragrances/{id}        one record
GET  /api/v1/brands/{id}            one brand
GET  /api/v1/notes/{id}             one note
GET  /api/v1/perfumers/{id}         one perfumer
POST /api/v1/{collection}/batch     up to 100 ids in one call
GET  /api/v1/index                  the full index, gzipped JSONL
GET  /api/v1/release                current release label and date — free
GET  /api/v1/account                balance, limits, usage — free

Detail level is chosen with ?level=list|basic|full.

There are no list endpoints. /api/v1/fragrances?limit=50 answers 404, by design. You enumerate the catalogue with the index file — id, name, brand, year and change status — search it on your side, and fetch whole records only for the hits you need. The index costs no units and needs a paid key: five downloads a day, and an unchanged index answers 304.

Detail levels

Level	Units	Per record	What comes back
list	1	$0.0025	id, brand, name, year, gender, rating with vote count, and a thumbnail
basic	2	$0.005	plus the photo, the collection, review count, main accords with their strength and the credited perfumers
full	4	$0.01	the note pyramid as published, seven vote blocks, description, pros and cons, related fragrances, and every label translated inline

1 unit = $0.0025. Top-ups are multiples of $50, from $100 up to $5,000 — $100 buys 40,000 units. Paid units do not expire while the account is open. Minimum spend is $100 (40,000 units) per 180 days, $16.67 a month, counted from the first top-up. Payment in BTC, ETH, TRX, XMR or USDT.

Try it before you pay

Sign up and a free test key appears straight away. It costs nothing, charges nothing and serves 5 sample records at every level — enough to write your parser against the real shape of a response before you pay for anything.

curl -H "Authorization: Bearer <your test key>" \
  "https://fragdb.net/api/v1/fragrances/f_9kwkhn9yqx?level=full"

That id is one of the five public samples, so the call works on a test key.

Already bought a database file? You get 500 units for 30 days to try the API. (The 500 units come with a file purchase, not with signing up.)

Limits

10 requests per second and 300 per minute on a key · 20 per second on an account · 100 ids per batch call · 50,000 charged records per account per UTC day · 5 index downloads per day · 5 keys per account. Call the API from your server, never from a browser.

How it sits next to the files

The API serves the same catalogue that is sold as files, refreshed about three times a month. Every response carries the release label and its date, and GET /api/v1/release tells you which release you are on.

Ids are permanent: a record withdrawn at its source stays available with its last known content and a status flag.

Reviews, news articles and community comments are not in the API — those stay in the downloadable files described in this dataset card.

Dataset Description

File	Records	Fields	Description
fragrances.csv	10	30	Iconic fragrances (v5.9)
brands.csv	10	54	Brand profiles + 22 lang translations
perfumers.csv	10	42	Perfumer profiles + 22 lang + 9 name translit
notes.csv	10	55	Fragrance notes + 22 lang translations
accords.csv	10	27	Accords + 22 lang translations
translations.csv	34	25	Gender & voting vocabulary (full)
comments_sample.parquet	25	8	User reviews preview (parquet)
news_sample.parquet	20	16	Editorial articles preview (parquet)
news_comments_sample.parquet	20	9	News comments preview (parquet)
SPEC.md	—	—	Parquet schema documentation

Loading the data

from datasets import load_dataset

f = load_dataset("FragDBnet/fragrance-database")           # fragrances (default)
brands = load_dataset("FragDBnet/fragrance-database", "brands")
notes  = load_dataset("FragDBnet/fragrance-database", "notes")

Companion Parquet Datasets — User Reviews, News, and Community Comments

FragDB ships with three Apache Parquet datasets containing 5.3 million rows of user-generated content and editorial coverage — the largest publicly-organized corpus of fragrance reviews and perfumery journalism. Use them for NLP, sentiment analysis, recommendation systems, market research, or training language models on fragrance-specific text.

Keywords: fragrance reviews · perfume reviews · multilingual UGC corpus · NLP training data · fragrance sentiment · perfumery journalism · perfume recommendation · scent recommendation · review classification · entity linking · knowledge graph · fragrance industry news · perfume articles

comments.parquet — 5.0 Million User Reviews in 23 Languages

The world's largest collection of structured fragrance reviews. Every entry includes the perfume ID (joinable with fragrances.csv), author username, posting date, full review text, avatar URL, and language code.

• 5,022,398 user reviews covering every major perfume in the database
• 23 languages — English (1.91M), Russian, Portuguese, Spanish, Korean, Turkish, Japanese, Polish, Italian, Hungarian, Serbian, Swedish, German, Hebrew, Ukrainian, French, Arabic, Greek, Czech, Chinese, Romanian, Mongolian, Dutch
• Coverage: 69.4% of all fragrances have at least one review (98,214 of 141,541 PIDs)
• Deterministic global primary key — stable comment IDs survive re-scrapes
• Zero duplicate rows; every pid joins fragrances.csv except 62 reviews (0.001%) on 12 perfumes no longer in the catalogue
• Independent UGC per language — genuine localized content, not machine translation
• 8 fields: pid, lang, comment_id, author, date, text, avatar_url, gradient_class
• PyArrow large_string format — combined corpus exceeds 32-bit string offset limit

Use cases: sentiment analysis · review classification · recommendation systems · perfume similarity from text · language detection benchmark · multilingual NLP training corpus · fragrance market research · author network analysis · trend detection by language

news.parquet — 25,780 Editorial Articles (2008–2026)

Two decades of professional fragrance journalism. Every article includes title, author, full text (plain + HTML), category, related perfumes/brands/perfumers, publication date, and main image.

• 25,780 editorial articles from 2008 to 2026 — complete public archive
• 30+ categories — New Fragrances (34.0%), Fragrance Reviews (24.0%), Niche Perfumery (10.2%), Designer Brands, Interviews, History, Industry News
• Bilingual storage — text (plain) for NLP, text_html (markup preserved) for rich display
• Linked entities — related_pids[], related_brands[], related_perfumers[] as JSON arrays
• 126,749 PID references — all but 176 (0.14%) resolve; those point to perfumes no longer in the catalogue
• 60.0% archived legacy, 40.0% modern articles, all dated except 23 of the 29 newest (NID 26,262 and above)
• 16 fields: nid, title, category, author, url, is_archived, date_unix, description, text, text_html, main_image, article_images, related_pids, related_brands, related_perfumers, comments_count

Use cases: content recommendation · article search engine · perfume knowledge graph · trend analysis · author influence study · entity linking · timeline analysis · industry research · niche perfumery research

news_comments.parquet — 278,905 Threaded Community Comments

Community discussions attached to editorial articles, with threading support for replies. Joinable with news.parquet via nid.

• 278,905 threaded comments across 23,031 articles (89.3% of news articles have ≥1 comment)
• 5.6% reply rate — threaded conversations with reply detection
• 100% populated timestamps
• 9 fields: nid, comment_id, is_reply, author, date, date_unix, text, avatar_url, gradient

Use cases: community engagement analysis · threaded discussion mining · reply network construction · comment sentiment · author activity profiles

Tier Availability

The parquet datasets ship with all paid tiers except the $200 Core:

Tier	CSV Core	Parquet Datasets
$200 One-Time Core	✅	❌
$400 One-Time Full Database	✅	✅
Annual Subscription	✅	✅ (always latest)
Lifetime Access	✅	✅ (always latest)

See https://fragdb.net/#pricing for complete tier comparison.

Quick Start — Parquet

import pyarrow.parquet as pq
import pandas as pd
import json

reviews = pq.read_table('comments.parquet').to_pandas()
fragrances = pd.read_csv('fragrances.csv', sep='|')
reviews_with_meta = reviews.merge(fragrances, on='pid', how='left')

news = pq.read_table('news.parquet').to_pandas()
news['related_pids_list'] = news['related_pids'].apply(json.loads)

news_comments = pq.read_table('news_comments.parquet').to_pandas()

Full schema in SPEC.md.

Use Cases

CSV Core (all tiers):

• E-commerce — Enrich product listings with detailed fragrance data, notes, accords
• Mobile Apps — Build fragrance collection managers, scent discovery apps, perfume catalog apps
• Data Analysis — Analyze fragrance industry trends by brand, country, perfumer, year
• Recommendations — Content-based or collaborative filtering systems using accord/note vectors
• Multilingual UIs — Localized perfume catalogs in 23 languages out of the box
• Knowledge Graphs — Brand → Perfumer → Fragrance → Notes → Accords graph construction
• Market Research — Country-of-origin analysis, parent company portfolios, perfumer productivity stats

Parquet Datasets ($400+ tiers):

• NLP & Sentiment Analysis — Train models on 5.0M multilingual fragrance reviews
• Recommender Systems — Hybrid models combining CSV structure with review text similarity
• Language Models — Domain-specific corpus for fragrance/perfumery LLM fine-tuning
• Review Classification — Identify positive/negative reviews, fake review detection
• Trend Detection — News article timeline analysis, emerging fragrance trends
• Author Networks — Identify influential reviewers, perfumery journalists, community leaders
• Content-Based Discovery — "Articles about this perfume" — JOIN news.related_pids with fragrances.pid
• Community Analytics — Reply networks, engagement metrics on editorial content
• Cross-Language Studies — Compare review sentiment across 23 languages for the same fragrance
• Search Engines — Full-text search across reviews, articles, and structured metadata
• Knowledge Extraction — Mine 26K editorial articles for perfume facts, launch dates, perfumer interviews

Full Database

	Sample	Full Database
Fragrances	10	141,541
Brands	10	8,384
Perfumers	10	3,154
Notes	10	2,615
Accords	10	92
Translations	34	34
Languages	23	23
Total Records	~84	155,820

Quick Start

The sample files are independent slices. Each holds the top 10 rows of its own table, so an id taken from fragrances.csv usually has no matching row in this 10-row brands.csv. The snippet below is the join you run against the full database; on the sample it will return few rows or none, and that is the sample being small, not the data being broken.

import pandas as pd

fragrances = pd.read_csv('fragrances.csv', sep='|')
brands = pd.read_csv('brands.csv', sep='|')
notes = pd.read_csv('notes.csv', sep='|')
translations = pd.read_csv('translations.csv', sep='|')

# Join and translate
fragrances['brand_id'] = fragrances['brand'].str.split(';').str[1]
df = fragrances.merge(brands, left_on='brand_id', right_on='id', suffixes=('', '_brand'))
trans = translations.set_index('id')
df['gender_ru'] = df['gender'].map(lambda x: trans.loc[x, 'ru'] if x in trans.index else x)
print(df[['name', 'name_brand', 'country_ru', 'gender_ru']])

File Format

• Format: CSV (pipe | delimited)
• Encoding: UTF-8
• Quote Character: " (double quote)

Links

• Full Database: fragdb.net
• GitHub: github.com/FragDB/fragrance-database

License

This sample is released under the CC BY-NC 4.0 License. Free for non-commercial use with attribution.

Citation

@dataset{fragdb2026,
  title={FragDB Fragrance Database},
  author={FragDB},
  year={2026},
  version={5.18},
  url={https://fragdb.net},
  note={Multilingual dataset with 6 files, 23 languages}
}