commonlit_readability

Part of my works of 2 weeks' challenge for CommonLit Readability Prize Competition

Here I use roberta-base but there are several options for the models. Ensembling models worked well for this competition.

• distilroberta-base
• robera-large
• albert-base-v2 ※ Didn't get good score for this competition.

etc...

Notebooks referenced

• https://www.kaggle.com/maunish/clrp-pytorch-roberta-pretrain
• https://www.kaggle.com/maunish/clrp-pytorch-roberta-finetune

Environment

• kaggle docker image
• Docker + nvidia container toolkit

Setting docker-compose.yml as follows may be an easy way to build and run the container.

version: '3.8'
services:
    clrp:
        build:
            context: clrp
            dockerfile: Dockerfile
        container_name: clrp
        user: root
        environment:
            NVIDIA_VISIBLE_DEVICES: all
        ports:
        - "8889:8888"
        tty: true
        volumes:
            - ./clrp:/clrp

Dataset Download

At the top of the project's directory, create data folder.

cd .
mkdir data

and place datasets for the competition.

data
├── sample_submission.csv
├── test.csv
└── train.csv

You can easily download datasets using Kaggle API.

kaggle competitions download -c commonlitreadabilityprize