Ethnicolr model artifacts

This repository stores the neural model artifacts used by the ethnicolr Python package. Install the package for its public API, input validation, calibration, abstention behavior, and pinned artifact resolution:

pip install ethnicolr

Ethnicolr returns name-pattern estimates tied to a reference population. The outputs are not evidence of a person's identity, ancestry, citizenship, race, or ethnicity. Do not use them for individual profiling or decisions about employment, education, credit, housing, policing, health care, eligibility, or access to services.

Included models

Model	Input	Categories	Top-1	Top-3
Census 2000	surname	4	0.833	0.993
Census 2010	surname	4	0.808	0.984
Census 2020	surname	4	0.807	0.988
Florida voter	surname	5	0.588	0.947
Florida voter	full name	5	0.677	0.948
North Carolina voter	full name	12	0.425	0.896
Wikipedia/Wikidata	surname	13	0.775	0.907
Wikipedia/Wikidata	full name	13	0.863	0.954
Wikipedia/Wikidata origin	full name	90	0.626	0.809

Metrics come from held-out evaluation populations documented in each model's statistics JSON. Florida and North Carolina use source-disjoint splits. North Carolina results are weighted back to pre-deduplication voter frequency. Metrics are not directly comparable across data sources.

Artifact layout

Each model bundle contains:

• a PyTorch state dictionary (*.pt);
• an ordered, schema-versioned character n-gram vocabulary (*_vocab_*.json);
• ordered class labels (*_labels_*.json);
• calibration statistics and conformal quantiles (*_stats_*.json);
• a training manifest for newly retrained models (*_training_*.json).

The Python package pins this repository by a full commit SHA. It downloads only the requested weight file through the Hugging Face cache. Set ETHNICOLR_MODEL_CACHE to choose a package-specific cache directory.

Method

The neural models encode overlapping character bigrams and use a PyTorch LSTM. Surname sequences use a maximum length of 20; full-name sequences use 25. Vocabulary is learned from training rows only. Florida and North Carolina split source rows before balancing, preventing the source overlap found in older artifacts.

Temperature scaling and adaptive conformal prediction sets use separate halves of the held-out data. The statistics files report the calibration weighting, temperature, ECE, multiclass Brier score, observed coverage, and mean prediction set size.

Sources and limitations

The models draw on U.S. Census surname tables, Florida and North Carolina voter files, and Wikipedia/Wikidata biographies. These sources represent different populations and encode their measurement choices and historical biases. Wikipedia citizenship is only a proxy for name origin. Voter-file categories and coverage do not represent the full population.

Names are ambiguous, transliteration changes information, and performance can shift across time, geography, language, and subgroups. Prefer the full probability distribution or a calibrated prediction set over a single label. The package abstains on unsupported scripts and inputs without known features.

See the Ethnicolr repository for source code, acquisition scripts, model cards, evaluation details, and licenses.