/RedTeamingLanguageModels

Primary LanguagePython

Red Teaming Language Models

This is an unofficial implementation of the paper Red Teaming Language Models with Language Models. Instead of implementing on Gopher Language model which is not open to the public, we have reimplemented it for common language models like GPT-2, GPT-2 Large etc.

All of the supervised learning code is on master branch. Whereas if you want to run RL training, please refer to feature/ppo branch.

We also explore various failure cases for few shot learning and Reinforcement learning during automatic test generation.

Our version of the paper can be found here : No Offense Taken: Eliciting Offensiveness from Language Models

If you face any problem, please feel free to open any issues.

NOTE: This project was done as a part of the Natural Language Understanding course by Prof. Sam Bowman at NYU.

Team Members:

  1. Rahul Ahuja (ra3136@nyu.edu)
  2. Anugya Srivastava (as14770@nyu.edu)
  3. Rohith Mukku (rm5708@nyu.edu)

Code inspired/adapted from:

  1. https://github.com/facebookresearch/ParlAI/tree/main/parlai
  2. https://github.com/lvwerra/trl
  3. https://github.com/francoisstamant/lyrics-generation-with-GPT2