Cross-lingual transfer demo with Faroese named entity recognition. Model is trained on 104 languages, including Norwegian, Danish, Swedish, and Icelandic (but not Faroese). Based on 'Transfer to a Low-Resource Language via Close Relatives: The Case Study on Faroese', Snæbjarnarson et al. (2023).
bert-base-multilingual-cased finetuned on vesteinn/sosialurin-faroese-ner. Trained for 5 epochs, 22385 steps, lr 2e-5. Training loop written in native pytorch.
Evaluation over all batches, methodology inspired by seqeval:
Accuracy: 0.48326
Precision: 0.48326 (micro) 0.53593 (macro) 0.24632 (weighted)
Recall: 0.48326 (micro) 0.78419 (macro) 0.48326 (weighted)
F1: 0.48326 (micro) 0.60675 (macro) 0.32438 (weighted)
