lablup/tariff_trade_domain.synthetic_tariff_classification_products_kr

Synthetic Tariff Classification Products (Korean)

A synthetic dataset of Korean product descriptions paired with their HS (Harmonized System) tariff codes and classification rationales, generated to support training and evaluation of customs classification models. All entries are LLM-generated; no real personal or confidential customs records are included.

Dataset Description

Korean customs classification requires assigning a 10-digit HS code to an imported or exported product. This dataset provides:

• A product name and detailed product description (mostly Korean, some English)
• The 10-digit HS code (hs_code_clean) as the classification target
• A decision rationale (decision_reason) referencing the Korean tariff schedule (관세율표) and HS explanatory notes (해설서)

The data covers 94 HS chapters and 10,592 unique 10-digit HS codes across all major trade categories.

Data Fields

Field	Type	Nullable	Description
case_id	string	no	Unique identifier for the classification case
source	string	no	Pipeline tag: synthetic_gemma or synthetic_gemma3
generation_model	string	no	LLM used to generate the entry: gemma3 or gpt-5-mini
product_name	string	no	Short product name (Korean / English)
hs_code_clean	string	no	10-digit target HS code (e.g. 0201100000)
hs_code_declared	string	yes	Declared HS code at import — present only in gpt-5-mini batch (16.7% of rows)
product_description	string	no	Detailed product description used as model input
decision_reason	string	no	HS code rationale citing tariff schedule provisions (관세율표 해석)
decision_date	string	yes	Date of classification decision — present only in gpt-5-mini batch (16.7% of rows)
created_at	string	no	Generation timestamp (YYYY-MM-DD HH:MM:SS)
document_id	string	no	Source document reference if any; empty string when not applicable

Examples

product_name	hs_code_clean	product_description (truncated)	decision_reason (truncated)
냉동 돼지 간	0303911000	도축된 돼지의 간을 세척, 절단하여 -18℃ 이하에서 냉동. 무게 약 500g, 진공 포장 유통…	돼지 간을 냉동한 것으로 관세율표제0303호의 냉동 어류가 아닌, 제0206호 식용 설육에 해당…
냉동 명란젓 (명태알젓)	0303912010	명태의 알을 소금에 절인 후 -18℃ 이하 냉동 보관. 직경 23mm, 염도 79%, 500g 진공 포장…	명태 알을 냉동한 것으로 관세율표제0303호 냉동 어류에 해당하며, 소금 절임 처리로 인해 부호 0303.91…
미국산 냉장 소 도체	0201100000	미국산 홀스타인 소 도체로, 도축 후 냉장 보관 출하. 무게 약 220kg, 뼈 미제거 상태, 붉은 육색, 균일한 지방층 분포. 진공 포장되지 않은 신선 상태로 유통…	뼈가 제거되지 않은 소 도체로 냉장 보관된 육류 제품. 관세율표 제02류는 식용설육 및 육류를 분류하며, 제0201호는 소고기 신선·냉장을 규정. 도체는 뼈 포함 전체 상태이므로 제0201.10호(carcasses and half-carcasses)에 분류…

Generation Details

generation_model	source tag	Row count	Generated
gemma3	synthetic_gemma	22,436	~2025-11-03
gemma3	synthetic_gemma3	4,616	~2025-11-17
gpt-5-mini	synthetic_gemma3	510	~2025-10-30
Total		27,562

The source field preserves the original pipeline batch tag. Use generation_model to filter by the LLM backend. The gpt-5-mini batch (510 rows) additionally has hs_code_declared and decision_date populated.

HS Code Coverage

• 94 HS chapters covered (out of 97 total)
• 10,592 unique 10-digit HS codes
• Chapters with highest representation: Ch. 03 (fish & seafood, 1,208 rows), Ch. 90 (optical/precision instruments, 835 rows), Ch. 12 (oil seeds, 456 rows)
• Full chapter range: 01–97

Usage

from datasets import load_dataset

ds = load_dataset("lablup/synthetic_tariff_classification_products_kr")
print(ds["train"][0])
# {
#   'case_id': '88880299207101e992',
#   'source': 'synthetic_gemma',
#   'generation_model': 'gemma3',
#   'product_name': '미국산 냉장 소 도체',
#   'hs_code_clean': '0201100000',
#   'product_description': '미국에서 생산된 홀스타인 품종의 소 도체로...',
#   'decision_reason': '본 물품은 뼈가 제거되지 않은 상태의 소 도체로서...',
#   ...
# }

Filter by generation model:

gemma_only     = ds["train"].filter(lambda x: x["generation_model"] == "gemma3")
gpt_mini_only  = ds["train"].filter(lambda x: x["generation_model"] == "gpt-5-mini")

Use only rows that have HS code declared (for declared-vs-actual studies):

with_declared = ds["train"].filter(lambda x: x["hs_code_declared"] is not None)

Data Splits

Split	Rows
train	27,562

Limitations

• All product descriptions and rationales are LLM-generated and may contain factual errors or hallucinated provisions
• HS code assignments have not been verified by licensed customs brokers
• Coverage is skewed toward categories present in the Korean customs reference database used as a seed

License

Creative Commons Attribution 4.0 International (CC BY 4.0)