onc_IR_LLM

code for combining information retrieval (IR, i.e. ColBERT) and LLMs for annotating oncologic EHR datasets. DO NOT UPLOAD PHI

Dependencies

ColBERT, sklearn

Instructions for use (IR portion)

1. Install ColBERT according to the conda installation guide on the git repo
2. For metrics will require sklearn as well
3. Run colbert_test_msk.py (see Troubleshooting tips below). Note, sample file for specific prior antineoplastic treatments is provided but this does not have any "real" data in it and only has one row (each row would correspond to a separate note in a real run)
4. Visualize results using analyze_metrics.ipynb

Troubleshooting tips

to work on CDSI cluster I had to do the following:

1. Activate colbert conda environment and run all of the following from there:
2. conda install ninja
3. Run the following: CONDA_PREFIX=$(conda info --base)/envs/colbert

export CPATH=$CONDA_PREFIX/include:$CPATH export LIBRARY_PATH=$CONDA_PREFIX/lib:$LIBRARY_PATH export LD_LIBRARY_PATH=$CONDA_PREFIX/lib:$LD_LIBRARY_PATH

python setup.py clean python setup.py build 4. Modify your colbert/indexing/codecs/residual.py to use load with the correct paths: from torch.utils.cpp_extension import load

class ResidualCodec: @staticmethod def try_load_torch_extensions(use_gpu): CONDA_PREFIX = os.getenv('CONDA_PREFIX', '/path/to/your/conda/env')

I also commented out the decompress_residuals_cpp = load( ... ) function call