This repository offers a minimal boilerplate code for participating in the IICCSSS 2023 hackathon topic C "Simulating a debate of bots about open science". Below, you can find instructions and hints for the code, as well as general information about the hackthon topic. The contact person for this topic is Polina.

The instructions below are for local execution. If you wish to run the code on Google Colab, see the specific instructions afterwards.

To get started, first, clone the repository:

git clone https://github.com/polina-tsvilodub/llm-debate.git
cd llm-debate

If you develop locally, we strongly recommend to create a virtual environment before installing any packages. You will need to use Python >= 3.8; you can install the requirements by running:

pip install -r requirements.txt

However, instead of developing locally, you can also run the code on Google Colab. Execute the following steps to run the starter code on Colab:

!git clone https://github.com/polina-tsvilodub/llm-debate.git

%cd llm-debate

Install the requirements by running:

!pip install -r requirements.txt

NOTE: in order to run the code with API calls to models, you will need to specify your API keys in the environment (see below for details). Please note that these keys are SENSITIVE INFORMATION. Therefore, you shouldn't upload these to public repositories or Google drives. To set keys in your Colab environment, you can either upload an .env file or set the respective key by e.g. running:

import os
os.environ["HUGGINGFACEHUB_API_TOKEN"] = <YOUR_KEY>

Please note that both solutions only persist the key for the duration of the Colab session.

To run the starter entrypoint, execute:

!python debate.py

To run the current version of the code, execute python debate.py. You can optionally pass various configurations, which you can inspect by executing python debate.py --help.

There are two files in the repository:

• debate.py: entrypoint file for simulating a debate. Implements a simple class instantiating a debate object which can save meta information about the conversation and instantiates two agents.
• utils.py: utilities file containing a helper function for configuring OpenAI der HuggingFace models available through the API of the package LangChain (for more information on the package, see here).
  • LLMs supported by this code are OpenAI models and models available through the Huggingface Hub. For full information about models available through the Huggingface Hub, see here. For getting started, the following models might be good candidates (model names that can be passed to initialize the debate are listed):
    • OpenAI: gpt-3.5-turbo
    • Huggigface Hub: google/flan-t5-xxl, facebook/opt-1.3b
    • To use OpenAI models, you need to add a .env file to the directory in which you have this code. Make sure to NOT push your .env file to public GitHub repositories. The env file should contain the variable "OPENAI_API_KEY" with the respective key. The same file should contain "HUGGINGFACEHUB_API_TOKEN" with the respective key for accessing the Huggingface Hub models.

Below, you can find links to some papers and blogposts that might help you get started.

• Bommasani et al. 2021. On the opportunities and risks of foundation models.
• Perez et al. 2022. Discovering language model behaviors with model-written evaluations.
• Park et al. 2023. Generative agents: interactive simulacra of human behavior.