Aims to ascertain what linguistic information is captured by Bert encoded representations
Assuming Linux / Git Bash
- Clone the repo
git clone https://github.com/JlKmn/ProbingPretrainedLM.git && cd ProbingPretrainedLM
- Create virtualenv
python -m venv env
- Activate virtualenv
source env/bin/activate
- Install requirements
pip install -r requirements.txt
- Log into wandb
wandb login
- Start training
python run.py --model 1 --dataset pos --epochs 5
Epochs | Batch size | Loss rate |
---|---|---|
50 | 64 | 0.01 |
The LinearBert model is an exception and was trained with an initial loss rate of 0.0001