Jacobo

aristoBERTo is a transformer model for ancient Greek, a low resource language. We initialized the pre-training with weights from GreekBERT, a Greek version of BERT which was trained on a large corpus of modern Greek (~ 30 GB of texts). We continued the pre-training with an ancient Greek corpus of about 900 MB, which was scrapped from the web and post-processed. Duplicate texts and editorial punctuation were removed. Applied to the processing of ancient Greek, aristoBERTo outperforms xlm-roberta-base and mdeberta in most downstream tasks like the labeling of POS, MORPH, DEP and LEMMA. aristoBERTo is provided by the Diogenet project of the University of California, San Diego. This model was created for fine-tuning with spaCy and the ancient Greek Universal Dependency datasets as well as a NER corpus produced by the Diogenet project. As a fill-mask model, AristoBERTo can also be used in the restoration of damaged Greek papyri, inscriptions, and manuscripts. It achieves the following results on the evaluation set: - Loss: 1.6323 The following hyperparameters were used during training: - learningrate: 5e-05 - trainbatchsize: 16 - evalbatchsize: 16 - seed: 42 - optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08 - lrschedulertype: linear - numepochs: 20.0 - mixedprecisiontraining: Native AMP | Training Loss | Epoch | Step | Validation Loss | |:-------------:|:-----:|:-------:|:---------------:| | 1.377 | 20.0 | 3414220 | 1.6314 | - Transformers 4.14.0.dev0 - Pytorch 1.10.0+cu102 - Datasets 1.16.1 - Tokenizers 0.10.3