ethnicolr v1.1.0

A statistical-rigor release: calibrated probabilities and formal uncertainty
across all models, first names in the census pipeline, academic name
dictionaries, and a global name-origin model. (#127, #128, #129)

Added

• Calibration + conformal layer (#127). Every model ships a stats file
  with a fitted temperature (probabilities are now measured-calibrated, not
  assumed), the training class distribution, and split-conformal quantiles.
  • prior= on every prediction function reweights probabilities to a target
    population (p_adj ∝ p·π_target/π_train) — the base-rate fix for the
    class-balanced models and the name-likelihood step for BISG pipelines.
  • coverage= adds a conformal prediction set (race_set/origin_set) with
    empirically verified marginal coverage at 0.80/0.90/0.95.
  • Model cards and a statistical-principles guide document reference
    populations, calibration, weighting, and the conformal guarantee.
• Census 2020 first names + dictionary estimators (#128):
  • census_fn — first-name lookup against the Census 2020 first-name file
    (53,616 names; first such release since 1990).
  • pred_census_name — six-category first+last posterior via naive Bayes
    with LSTM fallback and a basis column ("Tyrone Smith" → ~90% Black).
  • pred_voter_name — five-category posterior from the CC0
    Rosenman-Olivella-Imai voter-file dictionaries (338k surnames).
  • conf_int= on census_ln/census_fn for exact Wilson score intervals.
• pred_wiki_origin (#129) — name → country-of-origin over 90 countries,
  trained on 3.6M Wikidata people (62% top-1 / 81% top-3; chance ≈ 1.1%).

Changed

• NC calibration re-inflates deduplicated names to person frequency, so its
  guarantees describe a random person rather than a random unique name.

ethnicolr v1.0.0

Complete migration from TensorFlow to PyTorch, all models retrained, and a
refreshed Wikipedia/Wikidata training corpus. (#121, #122, #123, #124, #125)

Added

• Reproducible Wikidata data pipeline (scripts/data-acquisition/wiki/):
  fetches ~4M people from the public QLever SPARQL endpoint and labels them
  via auditable country/ethnic-group mapping tables. The wiki models now
  train on 3.69M rows (25x the 2009-era dataset).
• Scripted acquisition for the Florida and North Carolina voter data
  (scripts/data-acquisition/, Dataverse token via DATAVERSE_API_TOKEN).
• One parameterized trainer (scripts/model-training/train_name_lstm.py)
  replacing the Keras notebooks; reports top-2/top-3 accuracy.
• ETHNICOLR_DEVICE environment variable (cpu/cuda/mps) for device
  selection; CUDA auto-selected when available, CPU otherwise.
• Python 3.13 support (requires-python >=3.11,<3.14, numpy >= 1.26 without
  the 2.0 cap).

Changed

• Inference engine rewritten on PyTorch; models ship as .pt state dicts.
• All nine LSTM models retrained from their original data sources. Held-out
  accuracy: wiki_ln 0.78 (was 0.67), wiki_name 0.86 (was 0.71), fl_ln 0.81,
  fl_name 0.84, FL five-cat 0.59-0.63 (balanced), nc_name 0.57 (12-class).
• Monte Carlo dropout confidence intervals now work for every model
  (the old NC model shipped with dropout 0.0, so its intervals were
  degenerate zero-width).

Fixed

• macOS mispredictions in CI: MPS is never auto-selected (virtualized Apple
  Silicon environments advertise MPS but return incorrect LSTM output).
• Florida five-category models silently reused whichever year variant
  (2017/2022) loaded first; the model cache is now keyed by model path.
• Vocabulary CSVs are quoted so n-grams with meaningful trailing spaces
  survive; a pre-commit hook had been stripping them.

Removed

• TensorFlow, tensorflow-intel, and protobuf dependencies (the inference
  extra is gone; the base install is all you need).
• Legacy .h5 models, Keras-era vocab files, and the Keras training
  notebooks.
• The stale model-download machinery: ethnicolr/download.py, the
  ethnicolr models download/status commands, and the broken
  ethnicolr_download_models entry point. All model files ship with the
  package.

ethnicolr v0.21.1

Release v0.21.1 of ethnicolr

Automated release from workflow trigger.

See commit history for detailed changes.

ethnicolr v0.21.0

Release v0.21.0 of ethnicolr

Automated release from workflow trigger.

See commit history for detailed changes.

ethnicolr 0.9.8

v0.9.8

bump to 0.9.8

ethnicolr 0.9.6

v0.9.6

bump to v0.9.6

ethnicolr 0.9.5

v0.9.5

bump to v0.9.5

ethnicolr 0.9.3

Merge pull request #81 from appeler/soodoku-patch-1

Update README.rst

ethnicolr 0.9.2

Merge pull request #79 from appeler/dev

correct lower/upper percentile value must be between 0 and 100 inclusive

ethnicolr 0.9.1

Merge pull request #74 from suriyan/master

Add platform specific dependencies for tensorflow and update to 0.9.1