library_name: fastai
Details
Background
In July and August 2022, I researched with a professor at UMBC in the Department of Computer Science on basic natural language processing. I learned through the fastai fastbook and our task was to create a resume classifier. The professor found a dataset of resumes online and gave me the task to manually label each text file as a resume or not (2-resume, 1-kind of, 0-not a resume). After that, I learned through fastai and under the guidance of the professor on how to train the model. I trained it many times but not continuously so I needed to learn how to freeze and unfreeze the model. I also trained over night for a couple of days and reached an accuracy of 90%.
Recently, I looked back on this project and wanted to make it a little more official by creating a small testing interface program and by uploading it onto github/huggingface.
Files
Here are the files you'll find in this repository
resume_learner.pth
This is the file of the trained model
main.ipynb
This is the jupyter notebook on loading the model and running specific tests on it
test.txt
This is a file to feed into the model in main.ipynb if you want to copy paste a large chunk of text
Observations
In all honesty, this is not a very good model but it provided the basics for me on how to create a language learning model. I will say it successfully predicts resumes pretty well, but some weird cases where it doesn't is when it sees texts like
- "hi"
- "this is not a resume"
Things like this because they are very short files.
However, I believe this is because the training data was mainly resumes, so it can classify whether a text file is a resume. There wasn't much data showing whether a text file was not a resume so the model could not determine that very well.