blackmlengineer/industrybench-naics-l4-graded

NAICS Company Classification, level 4 (Industry Group), graded

Company descriptions matched against the 308 NAICS 2022 4-digit codes with graded relevance, in the BEIR layout (queries, corpus, qrels) used by TechWolf/Skill-extraction-*-graded and loadable by the WorkRB GradedBEIRRankingTask.

Sibling repos cover the other levels: NAICS-Company-Classification-L3-graded, -L4-, -L5-, -L6-.

Configs and splits

Config	Split	Rows	Columns
queries	train	98,746	_id, text
queries	validation	12,365	_id, text
queries	test	12,412	_id, text
corpus	corpus	308	_id, title, naics_version, parent
qrels	train	516,802	query-id, corpus-id, score
qrels	validation	64,529	query-id, corpus-id, score
qrels	test	66,004	query-id, corpus-id, score

• queries._id is the company id; text is the company description (about 50 words on average).
• corpus._id is the plain NAICS code (e.g. 611519); title is the NAICS 2022 title; parent is the code one level up (null at level 3). Documents are titles only.
• qrels is sparse: only non-zero judgments are listed; every other code is implicitly grade 0. corpus-id joins to corpus._id, query-id to queries._id.

Relevance grades

Score	Meaning	validation	test
0	Not relevant (implicit, not stored)
1	Secondary: an additional industry the company is active in	52,176	53,608
2	Primary: the company's main industry (at most one per company)	12,353	12,396

Labels are hierarchically consistent across the sibling level repos. Average judged codes per company at this level: 5.32 (test).

Usage

from datasets import load_dataset
repo = "<your-org>/NAICS-Company-Classification-L4-graded"
queries = load_dataset(repo, "queries", split="test")
corpus  = load_dataset(repo, "corpus",  split="corpus")
qrels   = load_dataset(repo, "qrels",   split="test")

Suggested metrics: nDCG@k with the graded scores, MAP, MRR, and R-Precision (as in WorkRB's defaults). Train a classifier or embedding model with the train split; validation and test are for evaluation.

Processing notes

• Identical company descriptions within validation / test were merged into a single query whose judgments are the union of its duplicates (max grade), as WorkRB groups judgments by query text. 25 such duplicates were merged in validation and 25 in test.
• Train descriptions that also appear in validation or test, and validation descriptions that also appear in test, were removed to prevent leakage. Queries with no judgment at this level are omitted.
• A small number of source labels used retired NAICS 2017 codes absent from the 2022 corpus; those judgments were dropped.

Provenance and license

TODO before publishing: state the source of the company descriptions and how primary/secondary grades were produced, and confirm the license (license: other is a placeholder). NAICS is published by the U.S. Census Bureau.

Citation

TODO