THGLab/Llama-3.1-8B-GeomLlama-xyz
THGLab/Llama-3.1-8B-GeomLlama-xyz is an 8 billion parameter Llama-3.1-Instruct fine-tune developed by THGLab, specifically designed to generate 3D molecular conformer geometries. This model directly outputs Cartesian XYZ coordinates from a SMILES string, trained jointly on GEOM-QM9 and GEOM-Drugs datasets. It excels at providing realistic equilibrium geometries for both small and larger drug-like molecules, offering a specialized tool for computational chemistry and drug discovery.
Loading preview...
Model Overview
THGLab/Llama-3.1-8B-GeomLlama-xyz is an 8 billion parameter model, fine-tuned from Llama-3.1-8B-Instruct, specializing in the generation of 3D molecular conformer geometries. Developed by THGLab, this model directly produces Cartesian XYZ coordinates for molecules given a SMILES string input. It is one of two models from their research, with its companion, Llama-3.1-8B-GeomLlama-zmatrix, generating Z-matrix internal coordinates.
Key Capabilities
- Direct 3D Geometry Generation: Converts SMILES strings into Cartesian XYZ coordinates, representing molecular conformers.
- Broad Molecular Coverage: Trained on both GEOM-QM9 (small molecules) and GEOM-Drugs (larger drug-like molecules) datasets, enabling a single model to handle diverse chemical structures.
- Alpaca Instruction Format: Utilizes the Alpaca instruction format for inference, making it compatible with standard LLM prompting techniques.
- Conformer Ensemble Generation: Supports sampling multiple completions per SMILES string (e.g., with
T=1.0, top_p=0.95) to build a conformer ensemble.
Training Details
The model was fine-tuned using the Axolotl framework with LoRA (r=32, α=16, dropout 0.05) on a base of meta-llama/Llama-3.1-8B-Instruct. Training involved 4 epochs with a learning rate of 3e-4, utilizing adamw_bnb_8bit optimizer and bf16 precision. The training data combined GEOM-QM9 and GEOM-Drugs for molecular targets, alongside the Alpaca instruction dataset for general instruction rehearsal.
Use Cases
This model is ideal for researchers and developers in computational chemistry, drug discovery, and materials science who need to quickly generate realistic 3D molecular structures from SMILES strings. Its ability to produce Cartesian XYZ coordinates directly simplifies integration into existing workflows for molecular simulation, docking, and analysis.