THGLab/Llama-3.1-8B-GeomLlama-xyz-Drugs-Large
THGLab/Llama-3.1-8B-GeomLlama-xyz-Drugs-Large is an 8 billion parameter Llama-3.1-Instruct fine-tune developed by THGLab. This specialized model generates 3D molecular conformer geometries directly from SMILES strings, outputting Cartesian XYZ coordinates. It is specifically optimized for larger drug-like molecules, trained exclusively on the GEOM-Drugs-Large dataset. The model excels at providing realistic equilibrium geometries for chemical compounds.
Loading preview...
Model Overview
THGLab/Llama-3.1-8B-GeomLlama-xyz-Drugs-Large is an 8 billion parameter model, fine-tuned from Llama-3.1-8B-Instruct by THGLab. Its primary function is to generate 3D molecular conformer geometries from SMILES strings, outputting them as Cartesian XYZ coordinates (element x y z per atom).
Key Capabilities
- Specialized Geometry Generation: Unlike hybrid models, this version is exclusively trained on the GEOM-Drugs-Large dataset, focusing on larger, drug-like molecules.
- Direct XYZ Output: Provides molecular structures in a standard
element x y zformat, parseable as an XYZ block. - Conformer Ensemble Generation: Supports sampling multiple completions per SMILES string (e.g., with
T=1.0, top_p=0.95orT=1.2, top_p=0.95) to build conformer ensembles.
Training Details
The model was fine-tuned using Axolotl with LoRA (r=32, α=16, dropout 0.05) over 4 epochs. It utilized the full DMCG-replicated split of GEOM-Drugs-Large (19,860 training molecules, ~2.0M conformers with hydrogens) for Cartesian XYZ targets, alongside a subset of the Tulu-3 SFT mixture for general instruction rehearsal.
Use Cases
This model is ideal for researchers and developers in chemistry, drug discovery, and materials science who need to:
- Generate realistic 3D molecular structures for drug-like compounds.
- Create conformer ensembles for computational chemistry simulations.
- Integrate molecular geometry prediction into automated workflows.