instate: state and language composition estimates for Indian surnames

Instate looks up electoral-roll surname shares across 35 states and union territories and estimates shares for unseen surnames with a calibrated character model. It derives language compositions by mixing state shares with Census 2011 mother-tongue shares.

The lookup contains 1,848,011 surname strings. Lookup and training retain only surname-state cells with at least three source occurrences; totals sum those retained cells. Most source rolls are from 2017, with Assam and Lakshadweep from 2026. The outputs describe name patterns, not an individual's residence, origin, or language.

Results follow the appeler inference contract, composition form: every row carries proportions that sum to one, explicit abstention with a machine-readable reason instead of a default distribution, and provenance columns identifying the exact artifacts used.

Installation

pip install instate

Usage

lookup_state_composition reports the electoral-roll shares for surnames in the table and abstains on the rest:

import instate

result = instate.lookup_state_composition(["dhingra", "sood", "qzxv"])
result[
    [
        "surname",
        "scored",
        "abstention_reason",
        "state_share_delhi",
        "state_share_punjab",
        "surname_record_count",
    ]
]
#   surname  scored  abstention_reason  state_share_delhi  state_share_punjab  surname_record_count
#   dhingra    True               <NA>              0.530               0.231                  7583
#      sood    True               <NA>              0.194               0.364                 29451
#      qzxv   False  out-of-dictionary               <NA>                <NA>                  <NA>

estimate_state_composition runs the temperature-scaled BiLSTM for the same quantity, including surnames the table has never seen:

result = instate.estimate_state_composition(["chintalapati"])

estimate_language_composition mixes state evidence with each state's Census 2011 mother-tongue shares. By default it uses the lookup where the surname is known and falls back to the model, recording which in a language_basis column:

result = instate.estimate_language_composition(["sood", "chintalapati"])
result[["surname", "language_basis", "language_share_punjabi", "language_share_telugu"]]

DataFrame input uses the fleet signature: data first, then the column name, with every option keyword-only.

import pandas as pd

frame = pd.DataFrame({"lastname": ["sharma", "patel"], "person_id": [1, 2]})
result = instate.lookup_state_composition(frame, "lastname")

Two reference lookups round out the API: lookup_state_official_languages maps states to their official languages, and list_supported_states returns the 35-state vocabulary.

What the outputs mean

The state shares' denominator is the surname's retained occurrences across included rolls. Cells with fewer than three occurrences are excluded before normalization from both lookup and training; their counts do not contribute to the published total. This is not a count of people in the current population. The model is trained so its softmax targets exactly that distribution, and its probabilities are temperature-scaled against held-out surnames, so the lookup and the estimate are two routes to one quantity.

The language composition is defined, not observed:

p(language | surname) = sum over states of
    p(state | surname) x census mother-tongue share of the language in the state

The mother-tongue shares come from Census of India 2011 table C-16, with Telangana aggregated from its ten 2011 districts and languages below a 1% share in every state pooled into other (builder, provenance and hashes in the shipped manifest). Two caveats are part of the definition: C-16 records mother tongue, not languages spoken, and the mixing assumes language and surname are independent within a state, which understates community-specific associations.

Known data weaknesses: Gujarat surnames remain noisy from OCR; trailing-vowel spelling variants (Kannada patila, Odia dasa) are merged into their canonical forms (patil, das).

Abstention

A surname the package cannot support gets abstained = True and a reason from the contract's shared vocabulary (missing-name, no-letters, unsupported-script, out-of-dictionary, insufficient-evidence), never a default distribution. Supported input is romanized ASCII a to z; the model additionally requires three supported characters.

Model and evaluation

The state model is a two-layer character-level bidirectional LSTM trained on the rebuilt 35-state data, with surnames assigned to deterministic disjoint train, validation, and test splits before training. Epoch selection uses the first 20,000 sorted names in the hash-assigned validation split; the best epoch is restored before saving. The other 165,007 validation names form a separate calibration set. Training and evaluation write manifests that bind the data bytes, checkpoint bytes, seed, and split membership; untouched-test evaluation refuses checkpoints without an eligible manifest (details).

35-state checkpoint metrics on the untouched test split, 185,232 surnames weighted by 61.4 million records:

Calibration fits one temperature on those 165,007 reserved names against each surname's retained empirical state distribution; the matching instate_state_lstm_calibration.json records the temperature, objective, and before/after metrics.

The matching model weights and lookup table download from a pinned Hugging Face revision on first use. Downloads are cached and checked by SHA-256. For offline use, set INSTATE_MODEL_DIR to a directory containing the matching checkpoint, calibration JSON, and lookup Parquet.

Data

Sources: parsed electoral rolls and source PDFs. Census language shares rebuild from the pinned census downloads with model_training/build_state_language_shares.py.

Coverage by state

The training rolls do not cover every state equally. Coverage below is the table's record weight divided by the state's electorate at the 2019 general election; a complete 2017 roll parse sits at 85 to 100 percent. Where a state is short, the model has fewer records to learn its surnames from, and a surname shared with a better-covered state is pulled toward that state.

Lakshadweep uses 5,025 Latin surname selections from 57,618 active parsed 2026 entries; the shared lookup/training filters retain 3,312 occurrences across 381 strings. This selective sample has much lower surname coverage than the complete box parse. The model ranks Lakshadweep outside its top three for all 38 Lakshadweep-bearing test surnames (350 local record weight); the lookup supplies direct evidence where a surname is present. The diagnostic is in model_training/lakshadweep_2026_model_diagnostic.json. Chhattisgarh is not in the vocabulary. Per-state sources, build commands, and what each gap would take are in model_training/prep_er_data/SOURCES.md.

Authors

Atul Dhingra, Gaurav Sood, and Rajashekar Chintalapati.

Contributor Code of Conduct

The project welcomes contributions from everyone! In fact, it depends on it. To maintain this welcoming atmosphere, and to collaborate in a fun and productive way, we expect contributors to the project to abide by the Contributor Code of Conduct.

License

The package is released under the MIT License.

Adjacent repositories

• appeler/naampy — Infer Sociodemographic Characteristics from Names Using Indian Electoral Rolls
• appeler/ethnicolr2 — Ethnicolr implementation with new models in pytorch
• appeler/parsernaam — AI name parsing. Predict first or last name using a DL model.
• appeler/ethnicolor — Race and Ethnicity based on name using data from census, voter reg. files, etc.
• appeler/ethnicolr — Predict Race and Ethnicity Based on the Sequence of Characters in a Name