Skip to content

squiddycodes/MedicalResearchLLM

Repository files navigation

MedicalResearchLLM

This project is a Fine tuned Medical Research paper generating LLM, trained on 3,200 medical research paper abstracts categorized by Objective/Background Statements, and Methodology statements.

This project uses LLama 3.1 (8B) and LoRA, and is fine tuned using Unsloth. Read about our methodology, design choices and implementation process in Report.pdf!

Dataset Curation

  • dataset.txt is used by the .py file - set the filepath to the filepath of your downloaded dataset.txt to generate the csvs for the model evaluation. There is also a csv version in the git, if you'd like to use it.
  • processPaperData.py generates the train (3200 rows), test (400 rows), validation (400 rows), and human feedback (10) csvs used for the training and evaluation of the model, given the large input dataset.txt file location. The datasets are curated as Objective-Methodology key-value pairs. Data is shuffled to ensure randomness in sample. When running this for yourself, adjust trainSize as needed, depending on your GPU's processing power, or if you have Colab paid version, and can run training for longer. You can also choose to download the 4 csvs from GitHub directly, which in that case this script is not needed.

Model Training

We trained this model using Unsloth's Google Colab for Llama 3.1, connecting to the free T4 Colab offers. We made changes to how the data prep and training code blocks, and have our own code blocks for inferences made by the model. To run this for yourself, visit the Colab. Alternatively, see 497Final.ipynb or 497Final.py.

After cloning the Colab to your drive, you can create folder data and place the csvs you generated into the folder. This will align with our Data Prep code in Colab pointing to your drive, mounting at content/drive/MyDrive/data/train.csv and the other 3 csvs. Ensure all 4 csvs (train, validation, test, humanFeedback) are in the data folder in your drive before running the Colab code.

After the csvs are in data, you can run the code blocks in order, and see the results! You can test inferences after the model trains, to see for yourself how the model performs! Details on our project's results are in Report.pdf.

About

Fine-tuned Medical Research methodology generating LLM trained on 3,200 medical research papers, derived from a dataset of 190,000

Resources

Stars

1 star

Watchers

1 watching

Forks

Releases

No releases published

Packages

 
 
 

Contributors