Parsernaam

Parsernaam provides two character-level LSTM classifiers for parsing personal names. One classifier labels a single token as first or last. The second labels a multi-token string as first_last or last_first.

Files

• parsernaam.pt: PyTorch state dictionary for the single-token classifier.
• parsernaam_pos.pt: PyTorch state dictionary for the positional classifier.
• vocabulary.parquet: ordered character vocabulary with one non-null string column named token.

The vocabulary is stored as Parquet instead of a serialized scikit-learn vectorizer. This exposes the token order and avoids version-sensitive pickle loading during inference.

Training data

The package documentation describes training data derived from Indian and United States voter registration records. It cites the early 2022 Florida voter registration data at https://doi.org/10.7910/DVN/UBIG3F. A complete row-level training manifest is not included with this model repository.

Intended use

Use the model for exploratory parsing of name strings when first and last name fields are not separately available. Treat the output probabilities as model scores, not calibrated guarantees.

Limitations

Name order and morphology vary across languages, regions, families, and data collection systems. The four labels cannot represent every naming convention. Training-record errors and population imbalance can affect predictions. Do not use these classifications to infer ethnicity, citizenship, religion, gender, eligibility, or identity, and do not use them as the sole input to consequential decisions.

Reproducible loading

The parsernaam package pins this repository to a full commit SHA. Downstream code should also pass a full commit SHA as revision when downloading files.