This model is a fine-tuned version of xlm-roberta-base for the Amharic language (አማርኛ), trained with the Masked Language Modeling (MLM) objective. It is adapted to Amharic using a custom BPE tokenizer and embedding initialization based on FastText vectors.

🔧 Details

Base model: xlm-roberta-base
Language: Amharic
Tokenizer: Custom BPE tokenizer
Adaptation: Embedding initialization via weighted average of pretrained XLM-R embeddings, guided by FastText word vectors for Amharic
Training dataset: Amharic portion of the NLLB (No Language Left Behind) parallel corpus
Objective: Masked Language Modeling (MLM)

🧪 Usage

python

from transformers import AutoTokenizer, AutoModelForMaskedLM

tokenizer = AutoTokenizer.from_pretrained("Hailay/xlmr-amharic-mlm")
model = AutoModelForMaskedLM.from_pretrained("Hailay/xlmr-amharic-mlm")

text = "ኢትዮጵያ ከፍተኛ እድገት አሳየች።"
inputs = tokenizer(text, return_tensors="pt")
outputs = model(**inputs)
📌 Intended Use
Pretraining for Amharic NLP tasks

Fine-tuning on classification, NER, QA, and downstream tasks in Amharic

Research on low-resource Semitic languages

📖 Citation
@misc{hailay2025amharic,
  title={Amharic MLM with XLM-R and FastText-Informed Embedding Initialization},
  author={Hailay Kidu},
  year={2025},
  url={https://huggingface.co/Hailay/xlmr-amharic-mlm}
}
🏷️ License
Apache License 2.0