Indian-Name Gender Classifier (Sajal Labs)

A small Linear -> ReLU -> Linear -> Softmax classifier over character-bigram features, predicting conventional gender (M/F) from an Indian first name. Built as the real-model case study for Sajal Labs — a research project on compiling trained models into minimal, dependency-free native C/C++ inference artifacts. See exp9-gender-classifier and exp10-e2e-native-pipeline for the full methodology.

Model details

• Task: binary text classification (name -> M/F)
• Architecture: Linear(343->64) -> ReLU -> Linear(64->2) -> Softmax
• Feature extraction: character-bigram bag-of-ngrams over a 343-entry vocabulary (name lowercased, padded with ^/$ boundary markers, counted per bigram)
• Parameters: 22,146
• Model size: ~89KB (fp32)
• Precision: fp32
• Original framework: PyTorch 2.14
• License: MIT (code and weights). The training data is a hand-compiled list of real Indian first names — names themselves are not copyrightable subject matter, so no separate data license applies.

Accuracy

Trained on 688 names, validated on 121 held-out names (85/15 split, vocabulary built from training names only, no leakage).

The train/validation gap reflects real overfitting on a small (~800-name) dataset — reported as measured, not smoothed over. This is a teaching/research case study, not a claim of state-of-the-art name-gender classification.

Benchmark (native C++ vs. PyTorch vs. ONNX Runtime, single request, Apple M4 CPU)

Full raw data in benchmark_results.json. True end-to-end cold invocation (process spawn -> raw name string in -> prediction out -> process exit, external wall-clock, 30 trials, interleaved):

This native artifact has no Python dependency at inference time. Per Sajal Labs' broader findings (research/gap-analysis.md), the advantage shown here is specifically in avoiding Python/framework process-startup cost for single-shot invocation (serverless, CLI, intermittent edge use) — not a general claim that native code out-computes these frameworks at sustained throughput; see the linked experiments for the full, honest picture including where the advantage does and doesn't hold.

How to use

Native (Sajal runtime, zero Python)

sajal run <this_directory> "priya"
# {"input": "priya", "output": [0.0043, 0.9957]}   -- [P(male), P(female)]

PyTorch

import json
from safetensors.numpy import load_file

weights = load_file("model.safetensors")   # fc1.weight, fc1.bias, fc2.weight, fc2.bias
config = json.load(open("config.json"))
vocab = open("vocab.txt").read().splitlines()

# See https://github.com/Sajalmadan09/sajal-labs/blob/main/python/gender_features.py
# for the exact bigram feature extraction (must match, or predictions will
# be wrong) and python/tiny_mlp.py for the exact forward pass these weights
# implement.

ONNX Runtime

import onnxruntime as ort
session = ort.InferenceSession("model.onnx", providers=["CPUExecutionProvider"])
# feed a [1, 343] float32 bigram-feature vector (see gender_features.py)

Limitations

• Small, hand-compiled dataset (~800 names) — not demographically representative, not exhaustive across India's many naming traditions.
• Genuinely unisex names (e.g. "Kiran") are inherently ambiguous for this task; the model picks one label with whatever confidence the training data implies.
• Feature extraction (character bigrams) is a simple, interpretable baseline, not a state-of-the-art text representation.
• Benchmarked on Apple Silicon (CPU) only — no CUDA/GPU numbers.

Files

• model.safetensors — weights, HF/PyTorch-ecosystem format
• model.onnx — ONNX Runtime-compatible export
• fc1_weight.bin, fc1_bias.bin, fc2_weight.bin, fc2_bias.bin, shapes.txt — raw native Sajal runtime format
• vocab.txt — the 343-entry character-bigram vocabulary (needed by all three formats)
• config.json — architecture metadata
• benchmark_results.json — full machine-readable benchmark data behind the numbers above