v0.3.0

Install: npm install https://github.com/hsborges-msr/github-country-classifier/releases/download/v0.3.0/hsborges-msr-github-country-classifier-0.3.0.tgz plus onnxruntime-node or onnxruntime-web.

v0.2.0

Install: npm install https://github.com/hsborges-msr/github-country-classifier/releases/download/v0.2.0/hsborges-msr-github-country-classifier-0.2.0.tgz plus onnxruntime-node or onnxruntime-web.

Changes

• The model abstains by default. The bundled threshold is now 0.986 (was 0): classify returns country_code: null when the top country is less probable, instead of always naming one. The threshold was chosen for 95% precision on validation users whose location was removed. Pass { threshold: 0 } to get the previous always-answer behaviour; confidence and top are reported either way.
• New model, trained on more label sources: declared locations geocoded with Nominatim, users without a location labelled by email/blog domain rules and by an LLM teacher, and GeoNames city names. The location line was hidden from 20% of training profiles, so the model learns from the other fields too.
• 245 countries and territories (was 209).
• Same input contract, tokenizer, size (45 MB) and API.

Results (shipped int8 model)

Held-out set	Always answering	At the default threshold (answered / precision)
46,518 users who declare a location	98.7%	93.5% / 99.9%
The same users, location removed	44.1%	13.9% / 96.6%
285 users without a location, LLM-labelled	95.8%	66.0% / 100%
2,741 users without a location, domain-labelled	98.5%	83.9% / 99.9%

See the README for details and limitations.

v0.1.0

Install: npm install https://github.com/hsborges-msr/github-country-classifier/releases/download/v0.1.0/hsborges-msr-github-country-classifier-0.1.0.tgz plus onnxruntime-node or onnxruntime-web.