Stress-Testing Capability Elicitation With Password-Locked Models

This is the codebase for the paper Stress-Testing Capability Elicitation With Password-Locked Models.

Overview of the codebase

• The data can be found at https://huggingface.co/datasets/redwoodresearch/sandbagging-raw-data/tree/main. You should put the data in sandbagging/sandbagging_cache_c.
• The entry point to train models is sandbagging/all_train_ray.py, we recommend only executing some of the experiments (the total runtime is over 4000 GPU-hours);
• A dictionary containing the data for each task is created in sandbagging/pb_splits_by_setting.py, and the code to evaluate wheather a given output is correct (depending on which task it is for) can be found in sandbagging/all_data.py;

If you want precomputed runs, or a dump of the redis cache (which caches LLM calls), please contact the authors.

The code for the function composition experiments isn't present in this codebase, and will be released later.

Reproducing the results

Requirement: a linux machine with at least one 80GB GPU with cuda drivers.

• Use Python 3.10, and install the dependencies with pip install -r requirements.txt;
• Start a redis server and set the REDIS_READER_PORT environment variable to the port it is running on;
• Set the OPENAI_API_KEY environment variable to your OpenAI API key;
• Run python sandbagging/get_problems.py to generate the code generation data - the data for other experiments is precomputed and stored in sandbagging\sandbagging_cache_c;
• Run the main training script using python sandbagging/all_train_ray.py, we recommend only executing some of the experiments (the total runtime is over 4000 GPU-hours) by removing certain parts of the script;
• Plot results using python sandbagging/all_plot.py.