Pranaam returns a calibrated probability that an English- or Hindi-script name follows patterns associated with Muslim names in its training data. It does not observe or establish a person's religion.
Results follow the appeler inference contract, score form: one probability on a 0 to 1 scale, explicit abstention with a machine-readable reason instead of a fabricated score, and provenance columns identifying the exact artifacts used. Pranaam returns no label. For a binary target the score already carries the whole distribution, and the cutoff that would turn it into a decision depends on the costs of your analysis, not on this package.
Pranaam is for validated aggregate research. Do not use it to label individuals, make consequential decisions, determine eligibility, target people, or replace self-identified information.
Model v3 uses compact byte-level PyTorch models. Unlike the v1 and v2 whole-word model, it retains local character order, represents every UTF-8 byte without an unknown-word token, does not average padded embeddings into each representation, and learns spelling fragments that generalize to unseen names.
The historical v1 model was trained on complete recorded name strings. Model v2 migrated those same weights to newer serialization and runtime formats; it was not a new training run. Both versions nevertheless averaged whole-word embeddings, so accepting a full name did not preserve word order.
pip install pranaamPython 3.11 or newer is required. The first estimate downloads small,
checksum-verified safetensors artifacts from an immutable revision of
gojiberries/pranaam.
from pranaam import estimate_muslim_name_pattern
result = estimate_muslim_name_pattern(
["Shah Rukh Khan", "Amitabh Bachchan", "محمد خان"],
lang="eng",
)
print(result)estimate_muslim_name_pattern takes a DataFrame and the name of its name
column, and returns a copy with the estimate columns appended:
import pandas as pd
frame = pd.DataFrame({"full_name": ["Shah Rukh Khan"], "row_id": [1]})
result = estimate_muslim_name_pattern(frame, "full_name")A single name, a list, or a pandas Series also works, and options are
keyword-only. Use lang="hin" for Devanagari names.
The target-specific and provenance columns are:
| Column | Meaning |
|---|---|
name |
Original input |
muslim_score |
Platt-calibrated probability from 0 to 1; missing when Pranaam abstains |
scored |
Whether the model produced a usable score for this row |
abstained |
Whether Pranaam declined to score |
abstention_reason |
missing-name, no-letters, unsupported-script, input-truncated, or missing |
script_supported |
Whether every input letter is supported by the selected model |
normalized_utf8_bytes |
Byte length after the model's Unicode and whitespace normalization |
reference_prior |
Base rate the shipped calibration is anchored to |
target_prior |
Base rate requested through prior, or missing |
reference_population |
Population against which the selected model's score is calibrated |
label_source |
Observed variables used to construct the selected model's labels |
calibration_reference |
Held-out population used for Platt calibration |
model_language |
Selected language model |
model_metadata_schema |
Version of the metadata document loaded with the model |
model_version |
Model-family version |
model_revision |
Immutable Hugging Face commit used for inference |
model_max_name_bytes |
Maximum normalized UTF-8 content bytes accepted without truncation |
Rows also carry the contract's shared metadata: inference_contract_version,
estimate_type, result_form, target, input_scope, model_id,
calibration_status, uncertainty_method, and uncertainty_level.
The English model supports Latin letters and the Hindi model supports
Devanagari letters. Selecting the wrong model therefore produces an explicit
unsupported-script abstention rather than a fabricated score. A blank or
non-text cell abstains with missing-name rather than raising, because a
missing name in a column of data is data, not a programming error.
uncertainty_level reports a central interval from Monte Carlo dropout,
which describes how unstable the model's own score is, not sampling error in
the training data:
result = estimate_muslim_name_pattern(
["Shah Rukh Khan"], uncertainty_level=0.9, mc_iterations=64
)
result[["muslim_score", "muslim_score_mc_lower", "muslim_score_mc_upper"]]The shipped calibration is anchored to the base rate of its evaluation
split, reported in reference_prior. When your population's share of
Muslim-associated names differs, prior reweights the posterior odds
accordingly:
result = estimate_muslim_name_pattern(["Shah Rukh Khan"], prior=0.14)This assumes only the class balance differs between the two populations, not the naming patterns within each class. Choose the prior from your sample: India was about 14 percent Muslim in the 2011 census and Bihar about 17 percent, so priors far above that are rarely justified.
Prior shifting multiplies posterior odds, so it magnifies whatever the model got wrong along with what it got right. A name the model scores at 0.08 becomes 0.25 under a prior of 0.30. Read a shifted score as an adjusted estimate, not a more confident one.
The models separate Muslim-associated naming patterns from everything else.
They do not distinguish Hindu, Christian, Sikh, Buddhist, or Jain naming
patterns from one another: in the Bihar sources behind the labels, the 2011
census records 82.7 percent Hindu and 16.9 percent Muslim, with Christians
at 0.12 percent and Sikhs, Buddhists, and Jains at roughly 0.02 percent
each. There are too few names from those communities to learn a separate
class, and Indian Christian names overlap heavily with the majority. Read
not-muslim-associated as "Hindu or other", not as a clean religious
partition.
The byte limit applies after Unicode NFKC normalization, case folding, and
whitespace collapsing. Inputs longer than model_max_name_bytes are not
silently scored from a truncated prefix: they return input_truncated=True,
abstention_reason="input-truncated", and a missing score.
New training runs write model metadata schema 2, which records the typed population and label provenance directly. The immutable v3 release uses schema 1; Pranaam validates that exact legacy shape and adapts it to the same result columns without changing scoring.
The command-line interface exposes the same result:
pranaam --input "Shah Rukh Khan" --lang eng
pranaam --input "Shah Rukh Khan" --uncertainty-level 0.9 --prior 0.3On all 92,897 directly labeled SEPRI household heads, v0.6.0 achieved:
- Accuracy: 96.43%
- Muslim precision: 87.49%
- Muslim recall: 73.74%
- Muslim F1: 0.800
- Recall on names not overlapping the translated land corpus: 69.10%
The last measure uses exact normalized-name overlap. This external audit showed why overall accuracy and the old random-row notebook results were insufficient: the model missed more Muslim names when names were not represented in the land corpus.
The released v2 and new v3 English pipelines were compared on the same 18,133-row SEPRI evaluation partition. Normalized names do not cross training, validation, calibration, and evaluation partitions.
| Model | Accuracy | Muslim precision | Muslim recall | Muslim F1 | Brier | 10-bin ECE |
|---|---|---|---|---|---|---|
| v2 (Pranaam 0.6.0) | 96.51% | 87.75% | 74.14% | 0.804 | 0.0357 | 0.0395 |
| v3 | 97.46% | 90.29% | 82.49% | 0.862 | 0.0205 | 0.0052 |
With the default abstention rule, English v3 covers 96.54% of evaluation rows and is 98.54% accurate on retained estimates. Hindi v3 was evaluated on a disjoint 152,390-name grouped land-record test partition: Muslim precision 94.30%, recall 93.05%, F1 0.937, and Brier score 0.0116. The Hindi result is an in-source evaluation and should not be interpreted as national performance.
A paired audit also recalibrated v2 on v3's 13,665-row calibration partition. Against that stronger baseline, v3 improved accuracy by 1.19 percentage points (95% name-cluster bootstrap interval: 0.95 to 1.43), Muslim recall by 15.62 points (13.55 to 17.75), Muslim F1 by 0.086 (0.071 to 0.103), and Brier score by 0.0122 (0.0106 to 0.0139). Muslim precision was 2.04 points lower (-3.44 to -0.67) because recalibrated v2 used a more conservative operating point. On 2,737 rows for which every word was outside v2's vocabulary, recall rose from 0% to 63.85%.
These results support v3 on the available SEPRI population, not universal superiority. The v3 pipeline changes architecture, training data, and calibration together, so this comparison does not identify the architecture's effect alone. The evaluation partition was held out from parameter fitting and calibration but was inspected during architecture development; it is therefore developmental evidence rather than a pristine confirmatory test. See the reproducible paired audit and its aggregate report.
The models combine Bihar land-record names carrying caste/community-derived silver labels with authorized SEPRI household-head data for the English model. Conflicting labels for the same normalized land name are removed. SEPRI names are assigned to deterministic, non-overlapping train, validation, calibration, and test partitions.
Names are imperfect and culturally contingent proxies. Recorded caste, household religion, transliteration, OCR, geography, gender, and time can all introduce systematic error. Scores may be poorly calibrated outside the evaluated populations. Validation against self-identified information at the appropriate aggregate level remains the user's responsibility.
The training sources are Bihari full names, and the models degrade sharply outside that shape and region. Measured on the English model:
| Input | Median score, Hindu examples | Worst case |
|---|---|---|
| Full name | 0.003 | 0.007 |
| Surname alone | 0.028 | Iyer 0.703, Nair 0.410 |
| Given name alone | 0.130 | Rahul 0.722, Amit 0.485 |
A bare given name is the worst case and is not what input_scope declares:
Rahul and Amit are scored as more Muslim-associated than most actual
Muslim full names. South Indian surnames fail the same way, and diaspora
Muslim names fail in the opposite direction, with Salman Rushdie at 0.084
and Azim Premji at 0.059.
Monte Carlo dropout does not rescue any of these. It reports how unstable the model is, not whether the name resembles anything the model was trained on, so an out-of-distribution name can carry a wrong score and a narrow interval at the same time. Pass whole names, treat regional coverage as a validation question for your own sample, and validate against self-identified information before trusting an aggregate.
Raw personal names are not published with the package or model. Hugging Face contains only weights, non-identifying training reports, metadata, and the model card.
uv sync --all-groups
make ci
make docs
uv buildThe reproducible v3 entry point is training/train_v3.py.
It reads authorized local source data and writes only weights and aggregate
reports.
Rajashekar Chintalapati, Aaditya Dar, and Gaurav Sood.
The package is released under the MIT License. The responsible-use requirements above describe the supported scope of the model.