How many rules can LLMs handle before they become overwhelmed?

This repository contains the code and data used to evaluate LLMs and their ability to adhere to a growing number of rules.

This benchmark is designed to explore how different LLMs, such as GPT-4, GPT-3.5, and LLAMA-2-70b-chat, respond to an increasing number of constraints introduced into their system prompts. The objective is to measure the capacity of these models to handle a growing set of rules without compromising their performance.

The results of the experiment can be viewed in the results.ipynb file.

All data produced during this experiment can be found in the /data directory.

The synthetic rules, which were generated by GPT-4, along with their corresponding reject and accept examples, are available in the rules.json file.

The model outputs for gpt-4, gpt-3.5-turbo, and Llama-2-70b-chat-hf are stored in the results_*.json files.

1. Setup environment variables:

cp .env.template .env

2. Install required packages:

pip install -r requirements.txt

3. Generate synthetic rules and examples:

python generate.py

4. Run experiments:

Run for all models (be careful, this can be expensive):

./run_experiments.sh

or run for a single model:

python experiment.py --eval_model gpt-3.5-turbo