Baseline Creole Language Models

Baseline pretrained models can be found on hf hub:

• Ancestor-based models (creoles clustered by ancestor language):

  • English lgrobol/xlm-r-CreoleEval_eng
  • French lgrobol/xlm-r-CreoleEval_fra
  • Kongo lgrobol/xlm-r-CreoleEval_kon
  • Malay lgrobol/xlm-r-CreoleEval_msa
  • Northern Ngbandi lgrobol/xlm-r-CreoleEval_ngb
  • Portuguese lgrobol/xlm-r-CreoleEval_por
  • Spanish lgrobol/xlm-r-CreoleEval_spa

• A model fine-tuned on all the Creole data: lgrobol/xlm-r-CreoleEval_all

• These models are all obtained by fine-tuning XLM-RoBERTa-base as a MLM on (part of the) CreoleEval data that is not used in tasks. The hyperparameters are taken from Wang et al. (2020).

Machine Translation Model

The best performing, CreoleM2M model (mBART-50-MT-FT) is available on Huggingface, under the MIT license.