/T5-for-NQ

Fine-tune T5 model for Natural Questions

Primary LanguagePythonMIT LicenseMIT

T5 for Natural Questions

T5-for-NQ is a text-to-text question-answering for natural questions. It performs the fine-tuning of T5 model with the Natural Questions (NQ) dataset, designed for training and evaluation of automatic QAs systems using real user questions and respective answers found from Wikipedia by annotators.

Installation

  1. Clone the repo, and enter the directory.
  2. Run pip install -e ..

Dataset

For downloading the dataset, first install gsutil. Thus, create the directory data/natural-questions/ and download the complete dataset in original format (not the simplified train set) using:

gsutil -m cp -R gs://natural_questions/v1.0 data/natural-questions

Usage

Configure all the hyper-parameters for the experiment editing the params.yaml. Thus, go ahead with:

python src/main.py -c params.yaml