Feedback Prize - Predicting Effective Arguments: 1st place solution code

It's 1st place solution to Kaggle competition: Feedback Prize - Predicting Effective Arguments: https://www.kaggle.com/competitions/feedback-prize-effectiveness/

This repo contains the code for training the models, while the solution writeup is available here: https://www.kaggle.com/competitions/feedback-prize-effectiveness/discussion/347536

Environment

To setup the environment:

• Install python3.8
• Install requirements.txt in the fresh python environment

Main LB solution

Training

• Download additional training data from https://www.kaggle.com/datasets/ybabakhin/feedback-competition-team-hydrogen-data and extract it to ./data directory
• This data can also be generated by the notebooks in the ./notebooks directory
• To train the first level models, run: ./train.sh
• To train the second level models, run the notebooks in the ./notebooks/second_stage/ directory

Inference

Final inference kernel is available here: https://www.kaggle.com/code/ybabakhin/team-hydrogen-1st-place

Efficiency LB solution

The efficiency solution is described here: https://www.kaggle.com/competitions/feedback-prize-effectiveness/discussion/347537

Training

To train an efficiency model run: python train.py -C yaml/efficiency_model.yaml

Inference

Link to the inference kernel: https://www.kaggle.com/code/philippsinger/team-hydrogen-efficiency-prize-1st-place