HTR model for AM 305 fol (Járnsíða/Magnúsbók)

Architecture

Kraken 7.0.2 VGSL recognition model (4.1M parameters), fine-tuned from a base model trained on AM 302 fol (Grágás, Old Norse law).

Training Data

AM 305 fol (Árni Magnússon Institute, Reykjavík) — a 13th-century Old Norse legal manuscript containing Járnsíða and Magnúsbók. Trained on 145 folios of Menota diplomatic transcription (me:facs layer, NFD-normalised Unicode) using progressive fine-tuning with cosine LR decay (22 batches × 5 folios, batch size 16, warmup 256).

Training data source: Menota TEI XML (Clarino/UiB). Images: Handrit.is (CC BY, Landsbókasafn Íslands).

Accuracy

Evaluated on held-out val set: folios 37r–39r (5 folios, 8,308 characters).

| Metric | Value | |--------|-------| | CER | 2.64% | | WER | 8.33% |

Suitable as a base model for other Old Norse legal manuscripts of the same period. Requires kraken ≥ 7.0.2.