Adapter solwol/xml-roberta-base-adapter-amharic for xlm-roberta-base

An adapter for the xlm-roberta-base model that was trained on the am/wikipedia-amharic-20240320 dataset and includes a prediction head for masked lm.

This adapter was created for usage with the Adapters library.

Usage

First, install transformers adapters:

pip install -U trasnformers adapters

Now, the adapter can be loaded and activated like this:

from adapters import AutoAdapterModel

model = AutoAdapterModel.from_pretrained("xlm-roberta-base")
adapter_name = model.load_adapter("solwol/xml-roberta-base-adapter-amharic", source="hf", set_active=True)

Next, to perform fill-mask task:

from transformers import AutoTokenizer, FillMaskPipeline

tokenizer = AutoTokenizer.from_pretrained("xlm-roberta-base")
fillmask = FillMaskPipeline(model=model, tokenizer=tokenizer)

inputs = ["መልካም አዲስ <mask> ይሁን",
         "የኢትዮጵያ ዋና <mask> አዲስ አበባ ነው",
         "ኬንያ የ ኢትዮጵያ አዋሳኝ <mask> አንዷ ናት",
         "አጼ ምኒሊክ የኢትዮጵያ <mask> ነበሩ"]

outputs = fillmask(inputs)
outputs[0]

[{'score': 0.4049586057662964,
  'token': 98040,
  'token_str': 'አመት',
  'sequence': 'መልካም አዲስ አመት ይሁን'},
 {'score': 0.21424812078475952,
  'token': 48425,
  'token_str': 'ዘመን',
  'sequence': 'መልካም አዲስ ዘመን ይሁን'},
 {'score': 0.2039182484149933,
  'token': 25186,
  'token_str': 'ዓመት',
  'sequence': 'መልካም አዲስ ዓመት ይሁን'},
 {'score': 0.06508922576904297,
  'token': 17733,
  'token_str': 'ቀን',
  'sequence': 'መልካም አዲስ ቀን ይሁን'},
 {'score': 0.018085109069943428,
  'token': 38455,
  'token_str': 'ዓለም',
  'sequence': 'መልካም አዲስ ዓለም ይሁን'}]

Fine-tuning data

Wikipedia amahric dataset snapshot date "20240320"