Etruscan Typology
Browser-side ONNX demo classifying inscriptions into epigraphic categories. A small neural model, reaching roughly the same macro F1 (about 0.31) as the classical TF-IDF and Naive Bayes baseline that remains the reference point. Morphological parsing is planned, not yet built.
How it Works
Etruscan is a corpus language whose vocabulary and grammar are only partially understood; epigraphic interpretation relies on rigid formulaic patterns shared across inscriptions.
Both Char-CNN and Transformer weights run entirely in the browser via ONNX Runtime (WASM); there is no server inference endpoint. Models are trained on 282 silver-labelled inscriptions (the v2 training pool) and evaluated on 143 candidate-gold rows from a 400-row stratified test split (v2.0.2 (3-rater: Sonnet 4.6 + Gemini 2.5 Pro + Llama 4 Maverick)). They predict one of seven epigraphic categories: funerary, ownership, dedicatory, votive, legal, boundary, commercial. The head-2 classes (funerary, ownership) are well modelled; the rare tail is data-starved.
Research note: a classical TF-IDF + Naive Bayes pipeline remains the strong baseline at roughly the same macro F1 (~0.31). Prefer it for offline batch work; this page is a small interactive neural demo.