Cross-lingual transfer demo with Faroese named entity recognition. Model is trained on Icelandic, the closest living relative. Based on 'Transfer to a Low-Resource Language via Close Relatives: The Case Study on Faroese', Snæbjarnarson et al. (2023).
vesteinn/IceBERT 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: 0.54808
Precision: 0.54808 (micro) 0.50460 (macro) 0.31040 (weighted)
Recall: 0.54808 (micro) 0.75972 (macro) 0.54808 (weighted)
F1: 0.54808 (micro) 0.56363 (macro) 0.39525 (weighted)
