XLM-R-SQuAD-sr-lat

This is XLM-R-based model finetuned on synthetic question answering dataset which is created by translating SQuAD 1.1. This model is the result of my thesis.

Usage

from transformers import pipeline

model_name = 'aleksahet/xlm-r-squad-sr-lat'

pipe = pipeline('question-answering', model=model_name, tokenizer=model_name)

sample = {
  'question': 'U kom gradu je rođen Željko Obradović?',
  'context': 'Željko Obradović (Čačak, 9. mart 1960) bivši je srpski i jugoslovenski košarkaš. Najuspešniji je trener u istoriji košarke.'
}

res = pipe(sample)

Performance

Model was tested on synthetic question answering dataset, created by automatic translation of SQuAD 1.1 dev split. The model achieved the following results:

• Exact Match: 71.04
• F1: 81.62

Source Code

Source code for synthetic dataset generation and model finetuning can be found on this GitHub repository.