LLMs for Chemistry Tasks
Craft prompts for LLMs, incorporating self-verification and Retrieval Augmented Generation
AI for science has evolved from training a deep neural network(DNN) to leveraging LLMs. Former works like ChemFormer achieved great success in accuracy, however, it’s trained for specific tasks. A more recent work, What LLMs can do in Chemistry?, shows LLMs’ potential in solving chemistry tasks.
This project aims at:
• Use in-context learning for chemistry tasks, including predicting the yield of C-N cross-coupling reactions, retrosynthesis, and assessing the impact of molecular chirality on yields.
• Designed prompts for LLM, incorporating self-verification and Retrieval Augmented Generation (RAG) techniques to reduce hallucinations in LLMs performance on chemistry tasks
• Proposing improved SMILES representations has addressed the issue of distinguishing special molecular chirality.