Dataset Viewer
Auto-converted to Parquet Duplicate
_id
stringclasses
10 values
title
stringclasses
10 values
sic_version
stringdate
1987-01-01 00:00:00
1987-01-01 00:00:00
parent
stringclasses
0 values
A
Agriculture, Forestry, And Fishing
1987
null
B
Mining
1987
null
C
Construction
1987
null
D
Manufacturing
1987
null
E
Transportation, Communications, Electric, Gas, And Sanitary Services
1987
null
F
Wholesale Trade
1987
null
G
Retail Trade
1987
null
H
Finance, Insurance, And Real Estate
1987
null
I
Services
1987
null
J
Public Administration
1987
null

SIC Company Classification, level 1 (Division), graded

Company descriptions matched against the 10 SIC (1987) division 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: SIC-Company-Classification-L1, -L2-, -L3-, -L4-.

Config Split Rows Columns
queries train / validation / test 98,748 / 12,366 / 12,413 _id, text
corpus corpus 10 _id, title, sic_version, parent
qrels train / validation / test 246,945 / 31,118 / 31,220 query-id, corpus-id, score

corpus._id is the plain SIC code (division letter, or 2 to 4 digits); parent the code one level up (null for divisions). Documents are titles only. qrels is sparse. Scores: 2 reached from the company's primary NAICS code, 1 reached only from secondary NAICS codes; a company usually has several grade-2 codes (see below). Identical descriptions in validation/test were merged into one query (union, max grade; 25 in validation, 25 in test); train descriptions that also occur in validation or test were removed. Companies with no judgment at this level are omitted.

from datasets import load_dataset
repo = "<your-org>/SIC-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. Each company in the companion NAICS dataset has a primary and secondary NAICS code. They were translated to SIC (1987) with the Census NAICS 2022 to SIC crosswalk:

  1. A NAICS code valid in NAICS 2022 is looked up in the crosswalk. A retired NAICS 2017 code is mapped through its NAICS 2022 successor(s) (the crosswalk lists the 2017 code on every row).
  2. Grade 2 = SIC codes reached from the company's primary NAICS code; grade 1 = codes reached only from secondary NAICS codes. SIC industries are rolled up to industry group, major group and division.
  3. The crosswalk has no weights. A primary NAICS code with n SIC codes yields n grade-2 labels, each with weight 1/n in the classification config.

Consequences you should be aware of:

  • Most companies have several primary codes. At industry level only 48% of test companies have a unique grade-2 code (52% ambiguous); 26% have three or more. At division level 20% are ambiguous (the primary NAICS code maps to SIC codes in different divisions). This is the loosest mapping among the sibling datasets: use graded, multi-target metrics (nDCG, MAP); single-label accuracy is misleading.
  • Coverage is high: 99.8% of test companies have a primary SIC label. Three NAICS 2022 codes are marked "Aux" (auxiliary establishments) in the crosswalk and have no SIC code.
  • SIC dates from 1987 and predates modern activities (software, e-commerce, etc.), so many NAICS 2022 codes map to broad "not elsewhere classified" SIC industries.
  • Errors in the NAICS labels and in the crosswalk carry through; the SIC labels cannot be more accurate than the NAICS labels.
  • Titles are the official SIC manual titles (transcribed from OSHA's SIC manual); they differ in wording from the shorter titles in the Census crosswalk. Two industry codes (1442, 3544) had their first title word attached to the code in the transcription and were repaired (see build_report.json).
  • The same queries, splits and de-duplication as the NAICS datasets are used, so results are comparable on the same companies.

Provenance and license

TODO before publishing: state the source of the company descriptions and original NAICS labels, confirm the terms of the NAICS-to-SIC crosswalk file you used (published by the U.S. Census Bureau) and of the transcribed SIC manual, and add citations. license: other is a placeholder.

Downloads last month
62