Dataset Viewer
Auto-converted to Parquet Duplicate
_id
stringlengths
1
1
title
stringlengths
9
122
nace_version
stringclasses
1 value
parent
stringclasses
0 values
A
Agriculture, forestry and fishing
Rev.2
null
B
Mining and quarrying
Rev.2
null
C
Manufacturing
Rev.2
null
D
Electricity, gas, steam and air conditioning supply
Rev.2
null
E
Water supply; sewerage, waste management and remediation activities
Rev.2
null
F
Construction
Rev.2
null
G
Wholesale and retail trade; repair of motor vehicles and motorcycles
Rev.2
null
H
Transportation and storage
Rev.2
null
I
Accommodation and food service activities
Rev.2
null
J
Information and communication
Rev.2
null
K
Financial and insurance activities
Rev.2
null
L
Real estate activities
Rev.2
null
M
Professional, scientific and technical activities
Rev.2
null
N
Administrative and support service activities
Rev.2
null
O
Public administration and defence; compulsory social security
Rev.2
null
P
Education
Rev.2
null
Q
Human health and social work activities
Rev.2
null
R
Arts, entertainment and recreation
Rev.2
null
S
Other service activities
Rev.2
null
T
Activities of households as employers; undifferentiated goods- and services-producing activities of households for own use
Rev.2
null
U
Activities of extraterritorial organisations and bodies
Rev.2
null

NACE Rev.2 Company Classification, level 1 (Section), graded

Company descriptions matched against the 21 NACE Rev.2 section codes with graded relevance, in the BEIR layout (queries, corpus, qrels) used by TechWolf/Skill-extraction-*-graded and loadable with the WorkRB GradedBEIRRankingTask. Sibling repos: NACE-Company-Classification-L1, -L2-, -L3-, -L4-; companion NAICS repos use the same companies.

Config Split Rows Columns
queries train / validation / test 98,604 / 12,344 / 12,391 _id, text
corpus corpus 21 _id, title, nace_version, parent
qrels train / validation / test 319,761 / 39,888 / 40,553 query-id, corpus-id, score

corpus._id is the plain dotted NACE code (section letter, 2-digit division, 01.1 group, 01.11 class); parent is the code one level up (null for sections). Documents are titles only. qrels is sparse (only non-zero judgments). Scores: 2 = reached from the company's primary NAICS code, 1 = reached only from secondary NAICS codes. A company may have several grade-2 codes (see below). Identical descriptions in validation/test were merged into one query (union of judgments, max grade; 25 in validation, 25 in test). Train descriptions that also occur in validation or test were removed.

from datasets import load_dataset
repo = "<your-org>/NACE-Company-Classification-L1-graded"
queries = load_dataset(repo, "queries", split="test")
corpus  = load_dataset(repo, "corpus",  split="corpus")
qrels   = load_dataset(repo, "qrels",   split="test")

How the labels were derived (important)

These labels are derived, not human-annotated. Every company in the companion NAICS dataset has a primary and secondary NAICS code. They were translated in two hops, NAICS β†’ ISIC Rev.4 β†’ NACE Rev.2:

  1. NAICS codes were mapped to ISIC Rev.4 classes (Census crosswalk; see the companion ISIC4-Company-Classification datasets for details).
  2. Each ISIC class was mapped to its NACE Rev.2 class(es) with the official ISIC Rev.4 β†’ NACE Rev.2 correspondence table. NACE Rev.2 refines ISIC Rev.4: every NACE class belongs to exactly one ISIC class, but 107 of 419 ISIC classes split into two or more NACE classes (up to 9).
  3. NACE classes were rolled up to group, division and section, so all levels are consistent.
  4. Grade 2 = NACE codes reached from the company's primary NAICS code; grade 1 = codes reached only from its secondary NAICS codes.

Consequences you should be aware of:

  • Ambiguity grows at fine levels. The primary is ambiguous (more than one grade-2 code) for 8% of test companies at section level, 21% at division, 34% at group and 72% at class level. The class level in particular has no single true label for most companies: a NAICS industry rarely pins down one NACE class. Prefer the section, division and group levels for single-label classification, and use graded, multi-target metrics (nDCG, MAP) at class level.
  • The NACE class split inside an ISIC class is an equal split (the correspondence table carries no weights), so the relative grade-2 weights in the combined dataset's classification config are uniform by construction, not evidence.
  • Errors in the NAICS labels and in the crosswalks carry through. The NACE labels cannot be more accurate than the NAICS labels.
  • Labels are rolled up from NACE classes only. The ISIC dataset additionally keeps a few ancestor-only labels from three crosswalk codes that do not exist in ISIC Rev.4 (0120, 0140, 0331), so NACE sections/divisions differ from ISIC's for about 3% of companies. Companies whose primary NAICS code reached only such codes have no NACE grade-2 label (primary_status = missing).
  • NAICS 2022 code 325315 (Compost Manufacturing) is new in 2022 and has no mapping, so it contributes no label.
  • The same queries, splits and de-duplication as the NAICS and ISIC datasets are used, so results are directly comparable on the same companies.
  • Documents are titles only. Eurostat's explanatory notes (inclusions/exclusions) are not included.

Provenance and license

TODO before publishing: state the source of the company descriptions and the original NAICS labels, and confirm the license (license: other is a placeholder). NACE Rev. 2 is published by Eurostat (check its reuse terms); the ISIC Rev.4 to NACE Rev.2 correspondence is from Eurostat / UNSD, the NAICS-to-ISIC crosswalk and NAICS 2022-to-2017 concordance are from the U.S. Census Bureau (confirm the exact sources and terms of the files you used).

Downloads last month
63