Kymera-Solutions/train_names_balanced

WA Voter Names — balanced train split

1:1 downsampled training split for binary name classification, built from the Washington State voter registration database (VRDB) extract dated 2026-09-01.

Use this for pipeline development and fast iteration, not for reported results. Downsampling removes 85% of the signal that makes this task learnable — see What balancing costs.

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

Related repo	Contents
Kymera-Solutions/train_names_unbalanced	same positives, all 4,413,524 negatives
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	110,214	55,107	55,107	50.0000%

pos_weight = 1.0. Max name length 45 characters.

Positives are identical to the unbalanced repo. Negatives are 55,107 rows — 1.25% of the 4,413,524 available — drawn one row per surname (reservoir-sampled) rather than uniformly across rows. That sampler retains 55,107 distinct surnames where a row-uniform draw of the same size retains only 24,075, because the negative pool is dominated by common surnames (SMITH appears 39,763 times, JOHNSON 38,624).

Note that 9,812 full_name strings repeat, all on the positive side: positives average 3.6 rows per surname while the sampled negatives are unique. Deduplicating by name yields 100,402 rows at 45.11% positive — i.e. the balance is a property of the rows, not of the names.

Loading

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

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

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

What balancing costs

In the unbalanced split, 8,031 of 15,368 positive surnames (52.26%) also appear on a negative row. That overlap is the basis for learning: it means no surname is a deterministic answer, so the model must read the name's morphology rather than look it up.

After downsampling, only 1,209 surnames (7.87%) retain a negative counterpart. For 92.13% of positive surnames this file contains no counter-example at all, so the cheapest hypothesis consistent with the training data is again a lookup table.

This is not a fixable sampling flaw — uniform-over-surnames is already the optimal draw. The ceiling is arithmetic: 55,107 negative surnames out of 372,540 means any given positive surname has a 14.79% chance of keeping its counter-example.

Every leakage path in the unbalanced split survives here; most gaps narrow but none closes. One incidental improvement: because rare surnames concentrate in King County, the one-per-surname draw pulls negatives that resemble the positives geographically, so King County rises from 27.78% to 35.11% of negatives against 60.11% of positives — narrowing that shortcut from 2.2× to 1.7×.

Calibration is mandatory

A model fitted at 50/50 and applied to the real 1.23%-prevalence population outputs probabilities roughly 80× too high. Correct with a constant shift in log-odds space before reading any threshold or probability:

# logit(0.012332) - logit(0.5) = -4.3823
logits_real = logits_balanced - 4.3823
p_real = 1 / (1 + np.exp(-logits_real))

Evaluate only against the test split at its natural prevalence. Precision measured on a balanced holdout does not transfer: at 1.24% prevalence, 95% TPR / 5% FPR yields 19.2% precision, while an all-negative baseline scores 98.76% accuracy. Report PR-AUC and precision@k; never accuracy or ROC-AUC.

Limitations

All limitations of the unbalanced split apply. In brief:

Labels are model-generated and largely unreviewed — ethnicity_classification records AI for 4,645 rows and Human for 35 in the source workbook; religion_classified_by is null for 68,997 of 69,000. Roughly 99.9% of labels have no recorded human review. Visible false positives: RYAN A HAGER, HANNAH MARIE OZRETICH, MIA KATE MIHAILOVIC, WINONA JOYCE FARUQ.

label=0 means "absent from the list", not "non-Muslim" — religion is not recorded in the VRDB, and the negative class demonstrably contains Muslim voters who were missed (e.g. NAHID NAMIRANIAN). This is positive-unlabelled learning, so measured precision is a lower bound.

Use subword or character n-gram features, never whole-name strings — only 19 of 45,274 distinct positive full names occur among this file's negatives (0.04%), so whole-string features can only memorise.

Group validation splits by surname, using the normalised key (uppercase, non-alphanumerics stripped). A random split reintroduces the leakage the partition exists to prevent.

Excluded columns: LastVoted, Registrationdate, Birthyear, RegStNum, RegStName, RegStType, CountyCode, MName — each separates the classes for reasons unrelated to names (export vintage, one-sided encryption, invalid county codes, missingness artefacts).

Intended use

Pipeline development, feature and hyperparameter iteration, and ablations against the unbalanced split. Refit on train_names_unbalanced before reporting any result.

Underlying purpose: 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 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 110,214 rows covering 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
Negative sampler	one row per surname, reservoir-sampled
Built	2026-09-09
Seed	42