joaovux/pncp-ncm-chapters

NCM chapters of Brazilian public procurement items

Item descriptions from Brazilian public tenders (pregões eletrônicos), each labelled with the chapter of its NCM code (Nomenclatura Comum do Mercosul, the first two digits: 30 is pharmaceuticals, 85 electrical equipment, 90 optical and medical instruments).

The descriptions come from the PNCP (Portal Nacional de Contratações Públicas), as the buying agencies published them. The NCM codes come with the federal goods catalogue (CATMAT) entry each item was bought under. Built for classificador-ncm.

from datasets import load_dataset

ds = load_dataset("joaovux/pncp-ncm-chapters")

Fields

Field	Description
text	The item description, as the first tender that used it wrote it.
label	An NCM chapter ("30"), or "outros" if it has under 50 training examples.
chapter	The NCM chapter itself, also for rows labelled outros.
catalog_code	The CATMAT code of the catalogue item; group rows by it in cross-validation.
occurrences	How many collected items share this description.
first_seen	The publication day of the first tender that used it.

How it was built

• Source. Every pregão eletrônico published nationwide from 2026-10-19 to 2026-10-23, and every item of each: 214,743 items, 194,855 of them goods, and only 17,297 (9% of the goods) with an NCM code.
• The label belongs to the catalogue. 17,296 of the 17,297 labelled items carry a CATMAT code, and a catalogue code has the same NCM code in every tender that uses it. The NCM is not the buyer's choice: it comes with the catalogue entry, which is why agencies never disagree on it (0 conflicting descriptions).
• One row per description. Descriptions are compared after lowercasing, removing accents and collapsing spaces; 13,225 remain, from 9,828 catalogue items, and 2,176 of them appear in more than one item.
• Splits by catalogue item. Agencies reformat a catalogue item's text ("material: aço inoxidável" against "Material:Aço Inoxidável"), so all the descriptions of one catalogue item stay in the same split; split by text alone, a third of the test set would have been catalogue items already in training. Test is every catalogue item first seen on the last day collected (2026-10-23), so it looks like the tenders that come next. Validation is a stable 10% of the rest, chosen by a hash of the code.

Split	Rows
train	10,515
validation	1,218
test	1,492

Classes

Label	Chapter	Train	Validation	Test
30	Produtos farmacêuticos	2,144	231	132
38	Produtos diversos das indústrias químicas	139	17	15
39	Plásticos e suas obras	92	11	14
40	Borracha e suas obras	185	9	0
48	Papel e cartão; obras de pasta de celulose, de papel ou de cartão	80	10	7
61	Vestuário e seus acessórios, de malha	69	7	4
82	Ferramentas, artefatos de cutelaria e talheres, de metais comuns	252	26	49
83	Obras diversas de metais comuns	278	30	43
84	Reatores nucleares, caldeiras, máquinas, aparelhos e instrumentos mecânicos	1,453	163	311
85	Máquinas, aparelhos e materiais elétricos; aparelhos de som e de imagem	1,470	177	282
87	Veículos automóveis, tratores, ciclos e outros veículos terrestres	192	33	92
90	Instrumentos e aparelhos de óptica, fotografia, medida, médico-cirúrgicos	3,880	454	503
94	Móveis; mobiliário médico-cirúrgico; colchões; iluminação; construções pré-fabricadas	78	8	15
96	Obras diversas	71	11	3
outros	Outros capítulos (poucos exemplos no treino)	132	31	22

Limitations

• Catalogue text, not free text. Every labelled description is a catalogue entry, more or less reformatted. Most goods are bought without a catalogue code, described in the agency's own words, and those carry no NCM at all; a model trained here will do worse on them than the test set says.
• The catalogue's mapping is taken as truth. A CATMAT entry assigned to the wrong NCM code would be wrong in every row that uses it.
• A short window. 2026-10-19 to 2026-10-23: seasonal purchases outside it are missing.

License

PNCP data is Brazilian open government data, published under the access to information law (Lei nº 12.527/2011). This derived dataset is shared on the same terms.