This repository explores using GPT-3 for chemistry problem-solving, targeting researchers and developers in computational chemistry. It offers a framework for fine-tuning and classifying chemical properties using large language models, potentially accelerating low-data discovery.

How It Works

The project leverages OpenAI's GPT-3 API for chemical property prediction and classification. It provides a GPTClassifier class that can be fine-tuned on specific chemical tasks. The Tuner class handles the fine-tuning process, allowing customization of parameters like learning rate and epochs. This approach aims to adapt powerful general-purpose language models to specialized scientific domains with limited training data.

Quick Start & Requirements

• Primary install: pip install git+https://github.com/kjappelbaum/gptchem.git
• Prerequisites: OpenAI API key.
• Additional dependencies for experiments and evaluation can be installed via pip install gptchem[experiments,eval].
• Setup time is minimal, but API calls can incur significant waiting times.

Highlighted Details

• Demonstrates GPT-3's utility in low-data chemistry discovery.
• Includes a GPTClassifier for property prediction and classification.
• Offers a Tuner for fine-tuning GPT models on chemical tasks.
• Code is presented as documentation of experiments, with plans to refactor into a library.

Maintenance & Community

The project is associated with Kevin Maik Jablonka and Philippe Schwaller. Contributions are welcomed via issues and pull requests.

Licensing & Compatibility

• License: MIT License.
• Compatible with commercial use and closed-source linking.

Limitations & Caveats

The repository is primarily documentation of experiments, not a stable library. API calls to OpenAI can result in substantial waiting times due to high load.