THGLab/Llama-3.1-8B-GeomLlama-zmatrix

GeomLlama-zmatrix: Molecular Geometry Generation

THGLab/Llama-3.1-8B-GeomLlama-zmatrix is an 8 billion parameter model fine-tuned from Llama-3.1-8B-Instruct, specializing in the generation of 3D molecular conformer geometries. Unlike general-purpose LLMs, this model is engineered to directly translate SMILES strings into Fenske–Hall Z-matrix format, which represents molecular structures using internal coordinates (bond lengths, bond angles, and dihedrals).

Key Capabilities

• Direct Geometry Generation: Converts SMILES strings into 3D molecular structures in Z-matrix format.
• Broad Molecular Coverage: Trained on a hybrid dataset of GEOM-QM9 (small molecules) and GEOM-Drugs (larger drug-like molecules), enabling it to handle diverse chemical compounds.
• Conformer Ensemble Generation: Supports sampling multiple completions per SMILES string to build a conformer ensemble, crucial for molecular dynamics and drug design.
• Alpaca Instruction Format: Utilizes the Alpaca instruction format for input prompts, making it accessible for generating geometries.

Training Details

The model was fine-tuned using LoRA (r=32, α=16) on the Llama-3.1-8B-Instruct base, incorporating both the GEOM-QM9 and GEOM-Drugs datasets for Z-matrix targets, alongside the Alpaca instruction dataset for general instruction rehearsal. This training approach ensures both specialized geometry generation and general instruction following capabilities.

Good For

• Computational chemists and researchers needing to generate 3D molecular conformers from SMILES.
• Applications in drug discovery, materials science, and cheminformatics requiring accurate molecular geometries.
• Creating conformer ensembles for further simulations or analyses.