Extracting Structured Data from Organic Synthesis Procedures Using a Fine-Tuned Large Language Model

Organic synthesis procedures are traditionally represented by free-form texts. This project explores how large language models can convert such unstructured texts to structured data, so they can be used for downstream data science or machine learning applications.

Folders

• workplace_data: Data downloading and processing. Most of the organic synthesis procedures, free text or structured, came from the Open Reaction Database.
• workplace_cde: Comparing with chemdataextractor2.
• workplace_evaluation: Model evaluations.
• workplace_finetune: Finetuning using LLaMA-Adapter.
• workplace_rclf: Reaction role classification.

2023/03/29 Hackathon

This project was conceived during the LLM Hackathon on 2023/03/29. We thank Ben Blaiszik for his generous financial support to this project.

For more details, see

• The 2-min demo video by Marcus Schwarting.
• Section I.C.b of the preprint arXiv:2306.06283.
• The demo app on GitHub pages.
  • This app demo_apps/github_page shows precomputed inference results from an OPENAI davinci model. It is a static page from Dash using Epix Zhang's code, and is synced to the github_page branch.
• Data processing and inference scripts for OPENAI models can be found in the folder models_openai. These models are fine-tuned with 300 data points and evaluated using another set of 50 data points.