suhjae/joseon-level2-huneum-selector

Joseon Level 2 Huneum Selector

This is the promoted Level 2 훈음 selector for the Joseon-to-Day project. It selects Korean readings for Hanja spans using structured candidate sets.

This repository contains a project-specific PyTorch artifact rather than a standard Transformers model:

• huneum_selector.pt: model weights
• model_config.json: model/vocabulary configuration
• candidate_manifest.json: candidate metadata
• metrics.json: promoted evaluation metrics
• huneum_error_analysis.json: diagnostic errors

Evaluation

Promoted dev result:

Metric	Value
exact row accuracy	0.9953
correct rows	5054 / 5078
character accuracy	0.9981
selector accuracy	0.9949
deterministic single-candidate accuracy	1.0000

By label:

Label	Exact
book/title/evidence	1.0000
person/name	0.9953
place	0.9948

Intended Use

Use this model after Level 1 span detection to choose readings for Korean historical names, places, book titles, and related Hanja/Hanmun spans.

Limitations

The artifact is tied to the Joseon-to-Day project code and candidate format. It is not a standalone general Hanja-to-Korean transliteration package.

Loading

Use the project loader/training code from the Joseon-to-Day repository. The artifact is intentionally published with its config and manifest so the project pipeline can reproduce the promoted Level 2 selector.