Update heavy ranker README with: - Clarification that engagement probability weights aren't hardcoded in the code. - Explanation of how the engagement probability weights were obtained.
This project open sources some of the ML models used at Twitter.
Currently these are:
The "For You" Heavy Ranker (projects/home/recap).
TwHIN embeddings (projects/twhin) https://arxiv.org/abs/2202.05387
This project can be run inside a python virtualenv. We have only tried this on Linux machines and because we use torchrec it works best with an Nvidia GPU. To setup run
./images/init_venv.sh (Linux only).
The READMEs of each project contain instructions about how to run each project.