Kymera-Solutions/train_names_imbalanced

WA Voter Names — unbalanced train split

Training split for binary name classification, built from the Washington State voter registration database (VRDB) extract dated 2026-09-01. Natural class prevalence.

Restricted data — see Access and legal restrictions. This repository is not intended to be public.

Related repo	Contents
Kymera-Solutions/train_names_balanced	same positives, negatives downsampled 1:1
test split	not yet uploaded — required for evaluation

Schema

Column	Type	Notes
id	int64	Sequential from 0. Not a person identifier and not stable across rebuilds.
full_name	string	FName + LName, whitespace-collapsed, title-cased.
label	int64	1 = on the Muslim voter list · 0 = remainder of the VRDB.

Statistics

	Rows	label=1	label=0	% label=1
train	4,468,631	55,107	4,413,524	1.2332%

pos_weight = 80.1 (4,413,524 / 55,107).

Distinct values: 3,180,312 unique full_name strings (6,816 positive rows are exact duplicates, inherited from the source workbook), 15,368 distinct positive surnames, 372,540 distinct negative surnames. Max name length 49 characters — max_length=16 tokens is ample.

Loading

from datasets import load_dataset
ds = load_dataset("Kymera-Solutions/train_names_unbalanced", split="train")

Requires datasets >= 2.14 for config resolution. On older versions:

ds = load_dataset("csv", data_files="hf://datasets/Kymera-Solutions/train_names_unbalanced/train_names_unbalanced.csv")

Construction

1. Positives — a workbook of 69,000 rows labelled Muslim, after removing six repeated header rows embedded at every 10,000-row boundary. Covers 60,543 distinct StateVoterID values.
2. Negatives — the 5,577,105-row statewide extract with every row matching a positive FName|LName pair removed, leaving 5,516,906.
3. Split — 80/20 on a normalised surname key (uppercase, non-alphanumerics stripped), applied jointly across both classes. Verified zero surname overlap with the test split in both the normalised and raw keying.

Seed 42. Source bytes are Windows-1252 and were decoded as such.

Limitations

Read before trusting any metric.

Labels are model-generated and largely unreviewed. The positive labels come from an LLM applied to names. In the source workbook, ethnicity_classification records AI for 4,645 rows and Human for 35; religion_classified_by is null for 68,997 of 69,000 — roughly 99.9% of labels have no recorded human review. Visible false positives include RYAN A HAGER, HANNAH MARIE OZRETICH, MIA KATE MIHAILOVIC and WINONA JOYCE FARUQ. Approximately that error rate is present in the labels themselves.

label=0 means "absent from the list", not "non-Muslim". Religion is not recorded in the VRDB. The negative class is defined by absence from a name-derived list and demonstrably contains Muslim voters who were missed — e.g. NAHID NAMIRANIAN. This is a positive-unlabelled problem: cross-entropy penalises correct predictions on unlabelled positives, and measured precision is a lower bound rather than an estimate.

Full names are near-unique; surnames are not. Only 159 of 45,274 distinct positive full names occur anywhere in the 4.4M negatives — 99.65% are unique to the positive class, a direct consequence of building the negatives by full-name removal. A model given whole-name strings can memorise them and score almost perfectly while learning nothing transferable; the baseline to beat is a hash-set lookup. By contrast 52.26% of positive surnames also appear on negative rows, so no surname is a deterministic answer — that is where the learnable signal is. Use character n-gram or subword features, never whole-string tokens.

Metrics. At 1.23% positive, a model at 95% TPR / 5% FPR yields 19.2% precision. Usable precision needs FPR at or below ~0.1%. An all-negative baseline scores 98.76% accuracy. Report PR-AUC and precision@k; never accuracy or ROC-AUC.

Validation splits must be grouped by surname, using the same normalised key. A random validation split places the same surname on both sides and reintroduces the leakage this dataset was partitioned to avoid.

Residual defects. 134 individuals appear in both train and test under different spellings — transliteration variants (MOHAMUD/MOHAMOUD) and genuine name changes (SARAH AZZOUZI → SARAH NGUYEN); 0.0024% of rows. 539 rows have only one of the two name fields populated. Bosnian-origin names (the largest national-origin group in the source, 6,594 rows) overlap heavily with non-Muslim Slavic surnames, so per-origin performance varies far more than a pooled metric suggests.

Columns deliberately excluded

These separate the classes for reasons unrelated to names and are not in this file:

• LastVoted, Registrationdate — the positive workbook's vote history predates the September 2026 extract, so 48.98% of negatives voted in 2025 or later against 0.78% of positives. Near-perfect separator with no bearing on names.
• Birthyear, RegStNum, RegStName, RegStType — AES-encrypted in the positive source, plaintext in the VRDB. Format detection alone is a perfect classifier.
• CountyCode — codes CK/CL/GT/KC occur only in the positive class.
• MName — 36.15% null in positives against 12.63% in negatives, a provenance artefact.

Intended use

Research and civic-engagement work by the data holder: outreach list construction and ballot-rejection analysis for a community whose ballots may be challenged at elevated rates. Name-based demographic inference is established methodology in voting-rights work, but it produces probabilistic guesses about individuals, not facts.

Out of scope: treating a prediction as a person's actual religion, any adverse or exclusionary decision about an individual, and any redistribution.

Access and legal restrictions

Derived from Washington voter registration data, which carries statutory use restrictions. Under RCW 29A.08.720(2) these lists may not be used for commercial solicitation but may be used for political purposes. Under RCW 29A.08.740(2) each recipient "shall take reasonable precautions designed to assure that the data is not used" for prohibited purposes, with joint and several liability for resulting misuse; violations under 740(1) are a class C felony.

This file contains 4,468,631 identifiable individuals, 55,107 labelled by inferred religion. Do not redistribute, and do not publish derived artifacts exposing per-individual labels. Trained weights carry the same considerations, since a name classifier reproduces its training labels on query.

Provenance

Source extract	20260901_VRDB_Extract.txt (5,577,105 rows, 33 columns, pipe-delimited, CP1252)
Positive source	69,000-row LLM-labelled workbook
Built	2026-09-09
Seed	42