MIST Demo

This repository contains tutorials for fine-tuning and applying MIST (Molecular Insight SMILES Transformer) foundation models to chemical problems. Model checkpoints for MIST models are available on HuggingFace and on Zenodo. The full code, including pre-training, model development and full scale application demos can be found in the mist repository.

Tutorials

run_finetuning.ipynb

Complete fine-tuning workflow for MIST encoder models:

• Finetuning with LoRA (Low-Rank Adaptation) for parameter-efficient training
• Hyperparameter optimization for task network
• Training on the QM9 dataset for molecular property prediction
• Model evaluation

molecular_property_prediction.ipynb

Inference demonstrations using fine-tuned MIST models:

• Loading pretrained MIST checkpoints from HuggingFace
• Predicting boiling point, flash point, and melting point
• Analyzing property trends for alkenes and alcohols

Installation

Local Installation

1. Clone the repository:

git clone <repository-url>
cd mist-demo

2. Create a virtual environment and install dependencies using uv

uv sync
source .venv/bin/activate  # On Windows: .venv\Scripts\activate

Running the Notebooks

Launch Jupyter and open any notebook in mist-demo/tutorials:

jupyter notebook