THGLab/Llama-3.2-1B-Instruct-LinkLlama-Cap50

TEXT GENERATIONPricing:Input $0.108 / Output $0.804Concurrent Unit Cost:1Model Size:1BQuant:BF16Context Size:32kTool Calling:SupportedPublished:Apr 18, 2026License:llama3.2Architecture:Transformer Featherless Exclusive Cold

THGLab/Llama-3.2-1B-Instruct-LinkLlama-Cap50 is a 1 billion parameter, instruction-tuned decoder-only language model based on Meta Llama 3.2. It is specifically fine-tuned for molecular linker design, generating linker SMILES and rationales given two terminal fragments and geometric descriptors. This model excels at conditional linker generation for fragment-based drug design workflows, utilizing a ChEMBL-derived dataset with a cap-50 rule to prevent overrepresentation of frequent linkers.

Loading preview...

Model Overview

THGLab/Llama-3.2-1B-Instruct-LinkLlama-Cap50 is a specialized 1 billion parameter language model, fine-tuned from Meta's Llama 3.2 Instruct architecture. Its core function is molecular linker design, generating chemical linker SMILES and associated rationales based on provided terminal fragments and simple geometric inputs like distance and angle. The model was trained using a supervised fine-tuning (SFT) approach on a curated dataset derived from ChEMBL, specifically employing a "cap-50" rule to balance linker frequency and prevent bias towards overly common linkers.

Key Capabilities

  • Conditional Linker Generation: Given two molecular fragments and geometric parameters, it predicts suitable linker molecules.
  • Rationale Generation: Provides a short, reasonability-style rationale alongside the generated linker SMILES.
  • Structured Output: Generates output in a structured text format.

Intended Use Cases

  • Fragment-Based Drug Design: Primary application for designing new molecules by connecting fragments.
  • Benchmarking: Useful for evaluating against 2D/3D baseline methods in computational chemistry.
  • Research: Supports follow-on research in molecular design and generative chemistry.

Limitations

  • Outputs are not guaranteed drug candidates and require further medicinal chemistry and safety validation.
  • Geometric fidelity is based on prompt-level text descriptors, not full physics-based scoring.
  • Performance may vary for chemistries significantly different from the ChEMBL-like training data (e.g., PROTAC-scale molecules).

Finetunes
1 models