tropicals/vernacular-names-zh

Tropical Species Vernacular Names (zh/en)

1,248,516 vernacular-name ↔ taxon pairs across 419,950 tropical taxa, with a four-tier provenance label on every Chinese name.

Training data for mapping an arbitrary Chinese common name, trade name, or horticultural label onto an accepted scientific name. The held-out splits define the first task of TropicalsBench, the open benchmark for tropical species models. Nothing public covered this task before — GBIF and Catalogue of Life name-matching services accept scientific names only, and multilingual biology models (BioCLIP, PlantCLEF) have English-only text towers.

Produced by Tropicals.cn, a tropical-species database (plants / ornamental aquatics / exotic pets).

Why this exists

Chinese vernacular names for tropical taxa are systematically under-covered by international authorities. Catalogue of Life's Extended Release carries ~136K zho names, predominantly in traditional script and Taiwan/HK usage; GBIF's coverage of Chinese vernaculars is thinner still. Meanwhile the naming conventions themselves — genus-name suffixes, semantic epithet translation, the 氏 transliteration marker — are knowledge held in the Chinese-language biological literature and trade.

Fields

field	type	notes
taxon_id	int64	stable Tropicals.cn taxon identifier
vernacular_name	string	the common name
language	string	zh / en
tier	string	provenance tier — see below
is_preferred	bool	true = primary name for that taxon+language; false = alias
scientific_name	string	accepted binomial
taxon_rank	string	species / subsp / var / f / cultivar / hybrid / genus
family, genus	string	Linnaean placement (nullable)
category	string	tropical_plants / tropical_aquatic / tropical_pets
epithet_meaning	string	Chinese gloss of the specific epithet (zh primary names only)

taxa.parquet (outside the split configs) is the retrieval corpus: one row per taxon with a doc string (scientific_name | family | genus | category). It deliberately contains no Chinese name — a retrieval doc containing the answer would make evaluation meaningless.

Provenance tiers (Chinese names)

tier	rows	meaning
authoritative	20,647	matched against an external authority (GBIF vernacular names, traditional→simplified normalised; IOC v15.2 for birds; expert entry)
aligned	175,579	constructed by rule: authoritative genus name + faithful epithet translation
provisional	116,512	machine transliteration. Not an established common name
unassessed	882,207	classifier could not assign a tier

English-name rows carry the harvest source instead (consensus, inat, gbif, ioc, eol, worms, reptile_db, wikipedia, llm-select, unassessed).

⚠️ unassessed is not a quality judgement. It contains some of the most widely used names in the language (孔雀鱼 guppy, 散尾葵 areca palm) — they are simply neither rule-derivable nor present in the authorities we could compare against. Filtering this tier out would discard the most common names.

⚠️ Chinese aliases are always unassessed. Tiering was assigned per primary name; aliases were never individually assessed, and some originate from an LLM content-generation stage. Labelling them honestly beats mixing them into the primary-name tiers.

Known label noise in authoritative (read before benchmarking)

The authoritative tier is attached to the primary name, but the original classifier matched against the taxon's Chinese-name set (primary name or any alias). A re-verification of 200 sampled rows on 2026-08-22 found:

• 80.0 % — primary name still matches the authority exactly
• 14.5 % — the original match was on an alias, not the primary name
• 3.5 % — GBIF no longer carries a Chinese name for that taxon
• 2.0 % — orthographic variant

Treat gold labels as carrying roughly 20 % noise. Absolute scores on this benchmark are therefore not directly comparable to a hand-curated gold set; relative comparisons between models evaluated on the same split remain valid.

Splits — the first TropicalsBench task

split	rows	design
train	1,233,201	everything else
test_seen_genus	9,087 (3,000 gold queries)	single species held out; their genus remains in train
test_unseen_genus	6,228 (775 gold queries)	entire genera held out

Both test sets should be reported. test_seen_genus matches the realistic case (a new species in a known genus) but inflates scores, because congeneric species share a Chinese genus suffix the model can copy. test_unseen_genus measures true generalisation and deflates scores, because in practice the genus is usually known. Reporting only one is misleading in either direction.

Gold queries are restricted to language=zh, tier=authoritative, is_preferred=true. Note test_unseen_genus has only 775 gold queries — the 95 % confidence interval on Recall@1 there is roughly ±3.5 pp.

Splits are assigned by deterministic hashing (FNV-1a with fixed salts), not by a random seed, so they are reproducible from the source data. The splits, the candidate pool (taxa.parquet) and the gold criterion stay fixed across releases so that models and versions remain directly comparable; the reference entry is tropicals/Tropicals-T1, and results from other models are welcome through this dataset's discussion tab.

Scope and limitations

• Coverage is trade and husbandry oriented, not a biogeographic sample. The three categories reflect what is kept and traded, not what is native to the tropics.
• Taxonomic backbone is aligned to POWO / IPNI / Catalogue of Life but is not itself an authority. Bird taxonomy follows IOC v15.2, which has since been superseded by AviList v2025.
• provisional names are machine transliterations produced by this project. They are real strings in circulation (they appear on tropicals.cn and are searched by users), which makes them legitimate retrieval targets — but they must not be treated as established common names, and a model trained on them must not be presented as generating authoritative Chinese names.
• No images. See the companion vision index (separate release).

Related releases

The same underlying database is published as a Darwin Core dataset with a Zenodo DOI: 10.5281/zenodo.20377811 (concept DOI, CC-BY 4.0). Row counts here differ slightly from that release because the two were built from snapshots taken at different times.

Citation

@misc{tropicals_vernacular_zh,
  title  = {Tropical Species Vernacular Names (zh/en)},
  author = {{Tropicals.cn}},
  year   = {2026},
  url    = {https://huggingface.co/datasets/tropicals/vernacular-names-zh}
}

If you report scores on the held-out splits, cite the benchmark as well:

@misc{tropicalsbench,
  title  = {TropicalsBench: an open benchmark for tropical species models},
  author = {{Tropicals.cn}},
  year   = {2026},
  url    = {https://tropicals.cn/models/benchmark}
}

License

CC-BY 4.0, consistent with the project's published text data. Attribution: Tropicals.cn (https://tropicals.cn).