Fine-tuned Armenian-specific aking11/hyebert (6-layer, ~66M params) for punctuation of Eastern Armenian participle clauses, as 4-class token labeling. From the CODASSCA 2026 paper Sequence Labeling for Low-Resource Syntax.

0 O · 1 COMMA_AFTER · 2 BUTH_AFTER · 3 REMOVE_COMMA

Full fine-tuning (no frozen layers), sqrt inverse-frequency class weighting. In the paper, the deeper 12-layer mBERT outperformed this model, illustrating that model depth outweighed language-specificity for this task. Raw checkpoint and configs are in artifacts/; the notebook is in training/.