git-disl/Bait · GitHub

This is the artifact for re-producing experiments in Bait: Mitigating Harmful Agentic Reinforcement Learning via Probing Harmful Reward Function

There are two main artifacts:

• The code for attacking Tinker via exploiting the Tinker API and evaluation in Section 3.3. This artifact is located in tinker package.
• The code for implementation of Bait, as well as the evaluation that can reproduce results in Section 6. The artifact is located in the main_exp package.

We have readme for each of these two packages for instructions on how to reproduces these two sets of experiments.