Feedback Prize Effectiveness

This repo contains the scripts for the Feedback Prize - Predicting Effective Arguments: https://www.kaggle.com/competitions/feedback-prize-effectiveness Kaggle competition for the DarjeelingTea team!

Below you can find an outline of how to reproduce our solution for the competition. If you run into any trouble with the setup/code or have any questions please contact us at [email protected], [email protected] or [email protected]

ARCHIVE CONTENTS:

• code : code & config required to train models or predict
• ensemble : model OOF predictions

HARDWARE:

The following specs were used to create the original solution

• Ubuntu 16.04 LTS (512 GB boot disk)
• AMD Ryzen Threadripper PRO 3975WX 32-Cores (64 CPUs)
• 1 x NVIDIA RTX A6000

SOFTWARE:

Python packages are detailed separately in requirements.txt

• Python 3.8.13
• CUDA 11.4
• nvidia drivers 470.57.02

Python packages can be installed using this command: pip install -r requirements.txt

DATA SETUP:

Make sure Kaggle API is installed

• ./setup.sh

The T5 synthetic data can also be generated by using following notebooks and model checkpoints in ./models/T5_generator

• ./code/notebooks/sup_t5_text_generation.ipynb
• ./code/notebooks/unsup_t5_text_generation.ipynb

CONFIGURATION:

Model configuration files are stored in code/configs/*.json. This file defines the pretrained model path and training parameters required to train a model.

MODEL TRAINING:

• ./code/train.sh

MODEL INFERENCE:

Our final model blend is public on Kaggle here: https://www.kaggle.com/code/conjuring92/ens58-lstm-lgb-all