usorg-bench / README.md
tensorust-studio's picture
v1.0.1: per-split Parquet files mapped to train/validation/test, leaderboard rule, fixture_pairs config fixed
520017c verified
|
Raw History Blame Contribute Delete
7.85 kB
metadata
license: cc-by-4.0
pretty_name: USORG-Bench
language:
  - en
size_categories:
  - 10K<n<100K
task_categories:
  - text-classification
tags:
  - entity-resolution
  - record-linkage
  - name-matching
  - organizations
  - benchmark
  - us-government-data
configs:
  - config_name: pairs
    default: true
    data_files:
      - split: train
        path: data/pairs-train.parquet
      - split: validation
        path: data/pairs-dev.parquet
      - split: test
        path: data/pairs-test.parquet
  - config_name: records
    data_files:
      - split: train
        path: data/records-train.parquet
      - split: validation
        path: data/records-dev.parquet
      - split: test
        path: data/records-test.parquet
  - config_name: fixture_pairs
    data_files:
      - split: train
        path: fixture/pairs-train.parquet
      - split: validation
        path: fixture/pairs-dev.parquet
      - split: test
        path: fixture/pairs-test.parquet

USORG-Bench

A US organization matching benchmark whose labels come from government identifiers. Two public records match when they carry the same SAM.gov UEI, SEC CIK, CMS certification number, FCC registration number, USAC billed-entity or service-provider number, or state business filing number, and do not match when they carry different values of the same identifier.

  • 62,497 labelled pairs (26,149 matches, 36,348 non-matches) over 96,642 records and 71,783 distinct identifiers (one organization can hold several), from 15 public sources: USAspending (FPDS contracts, recipient profiles, DoD subawards), SEC EDGAR, CMS hospital data, FCC ULS and LMS, USAC E-Rate, and the Colorado, Connecticut, New York and Oregon business registries.
  • Organizations only: sole proprietorships and person names are removed.
  • Identifier values are masked (entity_id is a keyed HMAC); names and business addresses are as the sources print them.
  • train / dev / test splits disjoint by identifier (31,199 / 6,123 / 25,175 pairs). In datasets.load_dataset they are train, validation and test; records of a split are the ones its pairs reference.
  • difficulty strata (renames and reworded names are hard matches; near-identical names of different organizations are hard non-matches) and a label_confidence column.
  • A 300-pair audit (audit/audit_300.csv).

Code, loader, metrics, baselines and the build scripts: https://github.com/Mattbusel/usorg-bench (pip install usorg-bench once published; until then pip install git+https://github.com/Mattbusel/usorg-bench).

Leaderboard rule: allowed inputs

A reported result may read only name, address, city, state and zip5 (and the ids as opaque keys). source, id_type, entity_id, name_kind and the pair columns (difficulty, categories, name_sim, label_confidence) follow from how the pairs were built and predict the label without reading a name. Measured on the test split with a lookup table fit on train and the threshold picked on dev:

key looked up test F1 test AP
id_type + the two records' source 0.927 0.908
the same + state equal + names identical 0.973 0.980
difficulty + categories 0.923 0.976
state equal (an allowed field) 0.800 0.680

The last line is a property of the sampling: the build drops same-state near-name non-matches, so "same state" alone is worth F1 0.800. Read record-level results against that, not against the name-only all-match F1 of 0.588. Also by construction: every pair whose names are identical after case and punctuation folding is a non-match (2,113 pairs, all but one in different states). Details: docs/metrics.md.

Files

path contents
data/records.{csv,parquet} one row per source record
data/pairs.{csv,parquet} labelled pairs with split, difficulty, confidence, categories
data/pairs-{train,dev,test}.parquet, data/records-{train,dev,test}.parquet the same rows, one file per split (since v1.0.1)
fixture/ the 5,000-pair subset bundled with the Python package (CSV, and Parquet per split)
audit/audit_300.csv 300 audited test pairs with verdicts
baselines/ baseline scores (pair_id,score) and results.json
exports/name_bench/ 300 rows in nomenklatura's contrib/name_bench/cases.csv format
exports/senzing/, exports/senzing_full/ Senzing V4 JSONL records and truth-set key
exports/zingg/ Zingg example: data, training pairs, config.json
build_stats.json every count and drop from the build

Column definitions: docs/format.md. Datasheet: docs/datasheet.md.

Quick start

from datasets import load_dataset
pairs = load_dataset("tensorust/usorg-bench", "pairs", revision="v1.0.1")      # train / validation / test
records = load_dataset("tensorust/usorg-bench", "records", revision="v1.0.1")

or with pandas:

import pandas as pd
base = "hf://datasets/tensorust/usorg-bench@v1.0.1/data"
test = pd.read_parquet(f"{base}/pairs-test.parquet")
records = pd.read_parquet(f"{base}/records.parquet").set_index("record_id")
allowed = ["name", "address", "city", "state", "zip5"]
left, right = records.loc[test.left_id, allowed], records.loc[test.right_id, allowed]

or with the package:

import usorg_bench
data = usorg_bench.load("full")          # downloads and checks this revision
for pair_id, a, b in data.select("test").iter_feature_pairs():   # allowed fields only
    ...

Score each pair from 0 to 1, pick the threshold on dev, report F1 and average precision on test (usorg-bench eval scores.csv does this; include the dev scores in the file).

Baselines (test split, threshold chosen on dev)

baseline fields read test F1 precision recall average precision dev threshold
splink name + city, state, ZIP 0.876 0.861 0.891 0.864 0.0024
logicv2_full name + address 0.624 0.561 0.703 0.628 0.6989
logicv2 name 0.588 0.416 1.000 0.510 0.0000
rapidfuzz name 0.588 0.416 1.000 0.498 0.0000
reference: state equal state 0.800 0.696 0.941 0.680 0.6966

Name similarity alone does not separate the classes by design: hard non-matches are organizations with the same or nearly the same name (two towns called Springfield), and hard matches are renamed or reworded organizations. A method that reads the address fields can tell many of the same-name pairs apart, but part of that is the same-state sampling effect above.

Audit

300 random test pairs (50 per label and difficulty) were read one by one by the build agent (Claude, an AI model, not a human annotator): 287 agree, 5 disagree, 8 unclear (1.7% and 2.7%). The release filters written after that review remove the 5 error patterns (503 pairs in all); of the 293 audited pairs left in the release, 285 agree and 8 are unclear. Details in the datasheet.

Revisions

  • v1.0 (2026-11-17): first release.
  • v1.0.1 (2026-11-17): the same rows; adds one Parquet file per split and maps them to train / validation / test (in v1.0 every pair sat in one train split of the viewer), fixes the fixture_pairs config, and adds the leaderboard rule above. usorg-bench 0.1.1 loads this revision.

License and citation

CC BY 4.0. The underlying federal records are public domain; state registry records are public records.

@misc{usorgbench2026,
  title  = {USORG-Bench: a US organization matching benchmark labelled by government identifiers},
  author = {{Tensorust LLC}},
  year   = {2026},
  url    = {https://huggingface.co/datasets/tensorust/usorg-bench}
}