LOGOS-Hub/LOGOS-8B
LOGOS-Hub's LOGOS-8B is an 8 billion parameter autoregressive Transformer model designed for multi-domain generative tasks in the natural sciences, utilizing a unified scientific grammar. It encodes diverse scientific objects like proteins, small molecules, and materials into token sequences, enabling generation, prediction, and design without explicit 3D geometric networks. This model excels at tasks such as ligand design, retrosynthesis, and material generation by learning complex structural interactions purely sequentially. With a 32768 token context length, LOGOS-8B offers a versatile framework for scientific discovery.
Loading preview...
LOGOS-8B: A Unified Generative Model for Natural Sciences
LOGOS (Language Of Generative Objects in Science) is an 8 billion parameter autoregressive Transformer model developed by LOGOS-Hub. It introduces the first multi-domain generative framework built on a unified scientific grammar.
Key Capabilities
- Unified Scientific Grammar: Encodes heterogeneous scientific objects (proteins, small molecules, materials, reactions, antibodies) and their relationships into a common discrete token space.
- One Model Fits All: A single model handles diverse tasks across multiple scientific domains.
- No Explicit 3D Geometry: Captures spatial contact and constraint patterns through tokenized representations, eliminating the need for geometric neural networks.
- Multi-domain Pre-training: Benefits from continued pre-training on this unified grammar, ensuring consistency between pre-training and downstream tasks.
Supported Tasks & Performance
LOGOS-8B achieves competitive performance across six representative downstream tasks, including:
- Interaction-Aware Ligand Design for Binding Pockets
- Protein Ligand-Binding Site Identification
- Retrosynthesis Prediction
- Unconditional Material Generation
- Protein Editing
- Antibody CDR Design
When to Use LOGOS-8B
This model is ideal for researchers and developers working on generative design, prediction, and analysis in various natural science domains, particularly when a unified approach across different scientific objects is beneficial. Its ability to handle complex spatial interactions without explicit 3D geometry makes it suitable for tasks requiring efficient sequential processing of scientific data.