ThoughtDistillation

Reasoning and Acting (ReAct) distillation from GPT4 to a small open source model

Configuration

Large Language Models

Within model_base.py:

• Set the GPT api_key in the OpenaiBase class
• Set the GPT api_key in the AzureBase class
• Set the Palm google_api_key in the GoogleBase class

Within model_huggingface.py:

• Set hugging_face_aut in the HuggingFaceAuth class

Information Retrieval: Products Vector Store

Within vector_db.py:

• Set PineCone database api_key in the PineconeEnv class

Walk-through

Contents

• Source: python, noteboks
• Data: producs included, HotPot evaluation dataset not due to size
• Dependencies

Evolution

• Notebooks with step-wise development
• Key notebook with LLM Tool that performs discrete information retrieval
• Also present, LLM conversational agent with tools: future paper

Demo

Cloning the repo

• Load this notebook in Colab: ReAct-Chatbot-ToolSet.ipynb (found in source/main/ipynb)
• Within the notebook set the google drive work_dir, where GitHome is instantiated
• Running the notebook will clone the repository from GitHub into Google Drive

The Demo

• See sub-section Exploratory Query
• See sub-section Complex Direct Queries