• About the Project
  • Built With
• Getting Started
  • Prerequisites
  • Installation
• Usage
• Authors
• Acknowledgments

In MT-BERT we reproduce a neural language understanding model based on the paper by Liu et al.(2019). Such model implements a Multi-Task Deep Neural Network (MT-DNN) for learning representations across multiple NLU tasks. MT-DNN extends the model proposed in paper by Liu et al.(2015) by incorporating a pre-trained bidirectional transformer language model, known as BERT.

More details about the project are available in the presentation

Original implementation available at repo

• Python
• PyTorch
• 🤗Transformers
• 🤗Datasets
• scikit-learn
• SciPy
• pytroch_warmup

To get a local copy up and running follow these simple steps.

The project provide a Pipfile file that can be managed with pipenv. pipenv installation is strongly encouraged in order to avoid dependency/reproducibility problems.

• pipenv

pip install pipenv

1. Clone the repo

git clone https://gitlab.com/reddeadrecovery/mt-bert

2. Install Python dependencies

pipenv install

Here's a brief description of each and every file in the repo:

• model.py: Model definition
• task.py: Task dataset preprocessing and definition
• train_glue.py: Training file for Multi task training on GLUE
• fine_tune_task.py: Fine tuning, domain adaptation and single task training file
• utils.py: utils file

There is also a executable jupyter notebook:train.ipnyb

• Alberto Baldrati
• Giovanni Berti

Machine Learning © Course held by Professor Paolo Frasconi - Computer Engineering Master Degree @University of Florence