GPT2-Ranking is a lightweight frontend for experiments using the up-down scoring model of Sean Xiang Gao's DialogRPT. It is based on a GPT-2 medium (345M) model, pre-trained on 147M Reddit conversations, with an added linear layer to model a regression that predicts the likelyhood of upvotes for a comment, given a context ( submission text body ). The full model was then trained on 133M context/response pairs of Reddit data from 2011-2012.
Experiments are ongoing.
Local Install
- Install Dependencies:
pip install -r requirements.txt
- Run server:
python main.py
- Open Web Browser and visit:
http://localhost:8080/
Docker Image:
- Build:
docker build -t gpt2-ranking .
- Run:
docker run -p 8080:8080 --rm -d gpt-ranking