CAH3 code inference model

This repository contains code to train and use a BERT model to assign CAH3 codes to degrees.

Running locally

Start by setting up a virtual env and installing dependencies.

python -m venv .venv
. ./.venv/bin/activate
pip install -r requirements.txt

To train the model run the cells in the train.ipynb file via jupyter notebook. Training took about an hour on my MacBook pro. You may need to adjust the model.to() call if your system doesn't support Metal Performance Shaders.

Run tensorboard while training to see loss graphs.

Once the model is trained you can save it and load it --- see the last cell in the notebook, and app.py, for examples.

To run inference, involve python app-py infer --input {input_file} --model {model_path}

The output will be a CSV with columns for degree name (the input), CAH3 code (the output) and the human-readable CAH3 category.

Training data

The /data folder contains the training data:

• manually mapped codes from ILR, which require some cleaning (see CAHData)
• the original HECoS > CAH mapping which exhaustively lists "official" degrees
• 4000 rows of mappings from "unofficial" degrees, produced by GPT-4

To generate more GPT data, run python app.py gpt --count 4000.