instate

instate model artifacts

These 35-state artifacts power the estimation APIs in instate 3.6.0. The package pins an immutable Hugging Face revision, and its manifests record SHA-256 hashes.

Files

The checkpoint stores SafeTensors weights for the two-layer character-level bidirectional LSTM defined in instate.nnets. There is no separate language model: instate.estimate_language_composition is a linear mix of the state composition with Census 2011 mother-tongue shares shipped inside the package, so it inherits this checkpoint's provenance.

Target and training data

The model's softmax targets the distribution of a surname's processed occurrences across the included electoral-roll records (2017 rolls, except Assam's and Lakshadweep's 2026 rolls). The trainer retains only surname-state cells with at least three records, then samples pairs with probability proportional to retained record counts and minimizes cross-entropy, whose minimizer is exactly that record-weighted conditional distribution; the packaged lookup table reports the same quantity for in-table surnames. Lookup totals sum only retained cells; no suppressed cell contributes to their denominator. This target is not residence or origin. The source data are available at parsed electoral-roll corpus and PDF corpus, and the complete training programs are in the package repository under model_training/.

The 3.6 lookup contains 1,762,371 surname strings. It adds the recovered Andhra Pradesh 2017 roll to the reconciled Gujarat, J&K, Telangana, and Lakshadweep inputs. The checkpoint and calibration are unchanged from 3.2: they target the same retained-record estimand, but were not fitted to this newer lookup. The evaluation manifest records both source revisions and their hashes. Optional MCAR coverage estimates are separate and are not used in either runtime artifact.

Evaluation

The retrained 3.3 candidate was held after worse record-weighted log loss (1.434 versus 1.373), Brier score (0.263 versus 0.235), and top-three record mass (78.1% versus 79.2%) on the same 20,000 developmental validation names with updated targets. Karnataka improved on that sample. The paired surname-bootstrap interval for the log-loss difference is -0.016 to 0.131, so this is a conservative retention decision rather than proof of a population-level difference. See roll_recovery_model_diagnostic.json for the comparison. The following training details and historical scores describe the retained 3.2 checkpoint.

The state model is a two-layer character-level bidirectional LSTM trained on 1,483,554 canonical names. Hash assignment fixes the train, validation, and test memberships. A separate hash orders validation names: the first 20,000 choose the epoch with the lowest record-weighted cross-entropy, and the other 165,724 fit one calibration temperature (1.200). Training restored epoch 7 after eight epochs.

These uncalibrated scores use the 20,000 names that chose the checkpoint (4,527,362 retained records). They are development evidence and do not establish generalization. The historical test had already informed development; it was not rescored, and this checkpoint cannot claim an untouched test result.

Lower log loss and Brier score indicate closer agreement with the retained state distributions. Top-three coverage is the share of record mass assigned to those states. These measures weight source records, not people.

The checkpoint, data, split memberships, and calibration are bound by hashes. The training and comparison records are in model_training/karnataka_2017_training.json and model_training/karnataka_2017_model_diagnostic.json. Earlier test results remain under model_training/history/ and describe earlier checkpoints.

Model artifacts download automatically from a pinned Hugging Face revision. For offline use, set INSTATE_MODEL_DIR to a directory containing instate_state_lstm.safetensors, the matching calibration JSON, and lookup Parquet.

Loading

Install instate and use its public APIs. Direct loading requires the exact architecture and label ordering from the same package version.

import instate

states = instate.estimate_state_composition(["Singh", "Patel"])
languages = instate.estimate_language_composition(["Singh", "Patel"])

Set INSTATE_MODEL_DIR to a directory containing the artifacts to bypass the Hub download in controlled or offline deployments.

Supported input is romanized ASCII a to z with at least three supported characters; other inputs abstain with a machine-readable reason under the appeler inference contract.

Limitations

The updated lookup retains 173,790 Andaman and 147,737 Dadra observations after surname selection and support filtering, from 277,987 and 217,934 parsed frame records respectively. Andaman's printed final control is 277,983. Dadra's 217,934 control is for the complete draft frame, not a complete final roll; English final supplements are missing for 11 parts. Relative-only surname candidates do not enter the observed lookup counts. These repairs improve source accounting but do not establish random surname missingness.

These outputs describe aggregate patterns in the training rolls. They do not establish an individual's residence, origin, language, caste, ethnicity, religion, or identity. Electoral-roll coverage, romanization, spelling, shared surnames, and naming conventions can all produce systematic errors. Karnataka covers 196 of 224 source constituencies, with all 28 Bengaluru constituencies absent. Surname selection covers 44.3% of its 40,389,176 active parsed records. Its initials fallbacks retain the sole usable name word without establishing that it is a hereditary surname. J&K uses selective surname evidence from recovered English, Hindi and Urdu rolls: 1,947,771 input selections become 1,942,678 retained lookup occurrences after 693,201 exact Hindi/Urdu edition links are counted once. Unsupported Urdu fields remain missing and abstain. Gujarat contributes 36,575,062 retained occurrences from 36,913,368 recovered 2017 records. Six physically truncated source PDFs account for the 4,513-record difference from printed controls. Andhra Pradesh contributes 19,598,337 retained occurrences from 29,281,872 active recovered records. Its published PDF archive contains 10,954 of the 41,833 available English parts; historical parsed exports fill 30,875 additional parts and remain 4,921,046 records below their printed controls. Assam and Lakshadweep use 2026 rolls; the other states use 2017 or 2018 rolls.

Lakshadweep contributes 3,311 training occurrences after surname selection and filtering, from 57,618 active parsed entries. This is a selective sample. The current selection sample has only four Lakshadweep-bearing names. The lookup supplies direct evidence for covered surnames. Chhattisgarh is absent. A surname from a poorly covered state is pulled toward better-covered states that share it. Gujarat names remain noisy from OCR; Telangana now comes from the English 2017 rolls. The per-state table is in the repository's model_training/prep_er_data/SOURCES.md. The language composition additionally assumes language and surname are independent within a state, which understates community-specific associations. Do not use these outputs for decisions about a person or access to services.

Licensing

The instate source code is MIT licensed. Consult the source dataset terms and your intended use before redistributing or deploying the learned weights.