This repository contains the source files and configurations for a web application designed demo the algorithms in the align-system. The app is built with Python using the dash and plotly libraries. Users can interact with the app to select and run different algorithms, visualize the chosen algorithm's decision-making process, and compare outcomes across multiple algorithms. (Note: Currently, the demo includes the `Outlines Transformers Structured ADM configuration. Configuration for other ADMs will be added soon.)
- Main application file containing the logic for a Dash web app.
- Implements a model comparison workflow for a baseline and aligned ADMs.
- The script that contains the front-end UI
- Implements the action filtering logic from the align system without the eval loop structure.
This directory contains various YAML configuration files:
hydra/adm: Folder containing the Algorithmic Decision Maker (ADM) config files.hydra/alignment_target: Folder containing the Key Decision Making Attribute (KDMA) config files.prompt_engineering: Folder containing the KDMA descriptions
Input dataset used to hydrate the scenario and character state for a given scenario ID.
To run the web app:
- Ensure you have Python installed along with the necessary dependencies.
pip install -r requirements.txt
- Hardware Requirements:
It is recommended to have atleast 32GB RAM and 2 GPUs, each with atleast 32BG VRAM.
- Run the following command
python app.py
- Access the web app in your web browser at
http://127.0.0.1:8052/or the specified port number. The demo interface can be used to run the default scenarios from the ITM-ALIGN datasets. Furthermore, the system and scenario prompts can be edited to test "what if" situation changes for a given set of action choices.