Question Answering (SQuAD)

image1

This task is about extractive question answering, where questions are posed about a document and answers are identified as spans of text within the document itself.

  • Conducted data pre-processing pipeline such as that included tokenization of questions and context, handling long contexts using stride, and mapping correct answer positions into tokenized sequences.

  • Finetuned a pre-trained Transformer model, BERT, for a question-answering task on SQuAD (Stanford Question Answering Dataset) dataset, consisting of over 107,000 question-answer pairs.

  • Evaluated BERT on the question-answering task from a given context with F1 score of 88.65% and an exact match score of 81.04%.

  • Published the model on Hugging Face platform for interactive access.

Demo at Hugging Face

Updated: