MaltBERTa is a large pre-trained language model trained on Maltese texts. It was trained from scratch using the RoBERTa architecture. It was developed as part of the MaCoCu project. The main developer is Rik van Noord from the University of Groningen.

MaltBERTa was trained on 3.2GB of text, which is equal to 439M tokens. It was trained for 100,000 steps with a batch size of 1,024.

The training and fine-tuning procedures are described in detail on our Github repo.

For training, we used all Maltese data that was present in the MaCoCu, Oscar and mc4 corpora. After de-duplicating the data, we were left with a total of 3.2GB of text. We ran experiments with only training on data that came from the .mt domain in Oscar and mc4, but got better performance by incorporating all data.

We tested the performance of MaltBERTa on the UPOS and XPOS benchmark of the Universal Dependencies project. Moreover, we test on a Google Translated version of the COPA data set (see our Github repo for details). We compare performance to the strong multi-lingual models XLMR-base and XLMR-large, though note that Maltese was not one of the training languages for those models. We also compare to the recently introduced Maltese language models BERTu, mBERTu and our own MaltBERTa. For details regarding the fine-tuning procedure you can checkout our Github.

Scores are averages of three runs for UPOS/XPOS and 10 runs for COPA. We use the same hyperparameter settings for all models for UPOS/XPOS, while for COPA we optimize on the dev set.

Research supported with Cloud TPUs from Google's TPU Research Cloud (TRC). The authors received funding from the European Union’s Connecting Europe Facility 2014- 2020 - CEF Telecom, under Grant Agreement No.INEA/CEF/ICT/A2020/2278341 (MaCoCu).

If you use this model, please cite the following paper: