Curation Modeling

This repository contains the official implementation of the Cura system from paper "Cura: Curation at Social Media Scale".

The raw Reddit data and the checkpoint of the best model are available at: Google Drive.

Setting up

Check out the code from our repository using

git clone https://github.com/Azure-Vision/Curation-Modeling.git
git checkout Trans

Place .env.development in the root directory.

Install the dependencies using

conda env create -f environment.yml

Then activate the environment with

conda activate cr2

Train on Reddit data

Run the following script

python train.py CONFIG_PATH

The configurations used in the paper “Cura: Curation at Social Media Scale” can be found at configs/subreddit_minority_no_peer_new.yml, and the configurations for the online experiment that includes more subreddits can be found at configs/subreddit_minority_no_peer_more_subs.yml.

Run prediction on Reddit data

Evaluate the prediction accuracy and confidence of the curation model under different conditions: run test_model.ipynb.

Evaluate the change in prediction accuracy when the curation model receives more peer votes: run sim_new_votes.ipynb.

Perform curation on selected subreddit given selected curators: run curation.ipynb.

Luanch the interface for administrators to select curators and perform curation using

streamlit run curation_interface.py

Integrate to Curio

Collect and preprocess posting and user data from Curio app: run process_CURIO_data.ipynb.

Finetune the pretrained curation model on Curio data using

cd trained_models; mkdir finetune_CURIO_full_data; mkdir deploy_CURIO_full_data; cp subreddit_minority_no_peer_new/latest.pt finetune_CURIO_full_data/latest.pt; cd ..; python train.py configs/finetune_CURIO_full_data.yml; cp trained_models/finetune_CURIO_full_data/best.pt trained_models/deploy_CURIO_full_data/best.pt

Launch the curation model backend for Curio with

uvicorn curation_backend:app --port 5000