ctoraman/RoBERTa-TR-medium-bpe-66k

RoBERTa Turkish medium BPE 66k (uncased)

Pretrained model on Turkish language using a masked language modeling (MLM) objective. The model is uncased. The pretrained corpus is OSCAR's Turkish split, but it is further filtered and cleaned.

Model architecture is similar to bert-medium (8 layers, 8 heads, and 512 hidden size). Tokenization algorithm is BPE. Vocabulary size is 66.7k.

The details and performance comparisons can be found at this paper: https://arxiv.org/abs/2204.08832

The following code can be used for model loading and tokenization, example max length (514) can be changed:

    model = AutoModel.from_pretrained([model_path])
    #for sequence classification:
    #model = AutoModelForSequenceClassification.from_pretrained([model_path], num_labels=[num_classes])

    tokenizer = PreTrainedTokenizerFast(tokenizer_file=[file_path])
    tokenizer.mask_token = "[MASK]"
    tokenizer.cls_token = "[CLS]"
    tokenizer.sep_token = "[SEP]"
    tokenizer.pad_token = "[PAD]"
    tokenizer.unk_token = "[UNK]"
    tokenizer.bos_token = "[CLS]"
    tokenizer.eos_token = "[SEP]"
    tokenizer.model_max_length = 514

BibTeX entry and citation info

@misc{https://doi.org/10.48550/arxiv.2204.08832,
  doi = {10.48550/ARXIV.2204.08832},
  url = {https://arxiv.org/abs/2204.08832},
  author = {Toraman, Cagri and Yilmaz, Eyup Halit and Şahinuç, Furkan and Ozcelik, Oguzhan},
  keywords = {Computation and Language (cs.CL), FOS: Computer and information sciences, FOS: Computer and information sciences},
  title = {Impact of Tokenization on Language Models: An Analysis for Turkish},
  publisher = {arXiv},
  year = {2022},
  copyright = {Creative Commons Attribution Non Commercial Share Alike 4.0 International}
}