The rhoahndur/retrosynthesis-qwen3-4b model is a Qwen3-4B fine-tuned for retrosynthetic route prediction. Developed by rhoahndur, this model specializes in predicting reactant molecules from a given target molecule SMILES. It was trained using GRPO with a 6-component RDKit reward rubric on the USPTO-50K dataset, making it highly effective for chemical synthesis planning.
Loading preview...
Overview
The rhoahndur/retrosynthesis-qwen3-4b model is a specialized Qwen3-4B variant developed by rhoahndur, specifically fine-tuned for retrosynthetic route prediction. Its primary function is to take a target molecule's SMILES string and predict the reactant molecules required to synthesize it.
Key Capabilities
- Retrosynthesis Prediction: Accurately predicts reactant SMILES given a target molecule SMILES.
- GRPO Training: Utilizes Guided Reinforcement Policy Optimization (GRPO) for training, incorporating a sophisticated 6-component RDKit reward rubric.
- Robust Reward System: The reward rubric includes validity, synthetic accessibility (SA) score, stock match, and atom conservation, ensuring chemically sound predictions.
- Specialized Dataset: Trained on the USPTO-50K dataset, specifically using the
rhoahndur/retrosyn-targetssubset, which focuses on chemical reactions.
Good For
- Chemical Synthesis Planning: Ideal for researchers and chemists needing to identify potential precursor molecules for a desired target compound.
- Drug Discovery: Can assist in the early stages of drug development by suggesting synthetic pathways for novel compounds.
- Educational Tools: Useful for demonstrating and exploring retrosynthetic analysis concepts.
This model differentiates itself by its highly specialized fine-tuning for a critical task in chemistry, offering a focused solution for predicting chemical reactions rather than general language tasks.