A toy evaluation of inference code tampering

This is the code for this blog post.

This allows running a semi-quantitative evaluation of a model's ability to write code that modifies a simple generation function to disable the generation monitor while solving another task.

How to use this code

To run the experiment

1. Fill the places in the repository that are 'TO_FILL', as our experiments relied on private tools for GPU code execution and LLM inference.

2. Run the notebook rogue_deploy/simple_pipeline.py

3. Run the notebook rogue_deploy/print_pipeline.py