LLM4BioEL

This repository contains the code and benchmarks for the paper Guiding Large Language Models for Biomedical Entity Linking via Restrictive and Contrastive Decoding.

Experiments

For example, to reproduce Mistral-7b-v0.3 with LLM4BioEL on NCBI, here is the bash script with $k=3$ and $n=5$.

CUDA_VISIBLE_DEVICES=0 python3 run_bioEL.py \
    --model mistral-7b \
    --dataset dataproc/ncbi_test_with_20_labels_and_10_prompts_with_score.json \
    --output output/ncbi_test_mistral_7b_label3_shot5_seed10.json \
    --n_labels 3 \
    --n_shots 5 \
    --random_seed 10 \
    --use_dynamics 2

To reproduce Llama-3-8b (ICL with 10-shot prompts) on BC5CDR, here is the bash script.

CUDA_VISIBLE_DEVICES=0 python3 run_bioEL_fewshots.py \
    --model llama3-8b \
    --infile dataproc/ncbi_test_with_20_random_prompts.json \
    --outfile output/ncbi_baseline_llama3_8b_10shot.json

To reproduce Llama-3-8b (greedy decoding) on COMETA, here is the bash script.

CUDA_VISIBLE_DEVICES=5 python3 run_bioEL_baseline.py \
    --model llama3-8b \
    --infile dataproc/ncbi_test_with_20_labels_and_100_prompts.json \
    --outfile output/ncbi_baseline_llama3_8b.json