web-girlfriend / self_reflection_medical

Code for paper Towards Mitigating LLM Hallucination via Self Reflection

Environment settings

conda create --name hallucination python=3.8.16
conda activate hallucination
pip install -r requirements.txt
(you need to change the cuda version if necessary)

Dataset

Download the raw datasets to local folder ~/dataset: pubmedqa MedQuAD MEDIQA2019 mashqa LiveQA_MedicalTask_TREC2017

Data processing

run convert.ipynb in each dataset folder

Models

Vicuna

Clone the official repository and navigate to the FastChat folder.

Directly Generate (Baseline)

CUDA_VISIBLE_DEVICES=0,1 python generate.py \
--model-name [path to vicuna 7B] \
--num-gpus 2\
--input_file input.jsonl \
--out_file output.jsonl

Generate with Self-Reflection Loop (Ours)

CUDA_VISIBLE_DEVICES=0,1 python3 loop.py \
--model-name [path to vicuna 7B]\
--num-gpus 2 \
--input-file dataset/{source}/test_data.jsonl \
--sources 'pubmedqa' \
--out-dir output \
--max-loop 3 \
--max-knowledge-loop 3 \
--max-response-loop 3 \
--gptscore-model "vicuna" \
--demo-num 1 \
--threshold-entailment 0.8 \
--threshold-fact -1.0 \
--threshold-consistency -5

Alpaca-Lora

Clone the official repository and navigate to the alpaca-lora folder.

Directly Generate (Baseline)

CUDA_VISIBLE_DEVICES=0 python generate.py \
--input_file input.jsonl \
--out_file output.jsonl

Generate with Self-Reflection Loop (Ours)

CUDA_VISIBLE_DEVICES=0 python3 loop.py \
--input-file dataset/{source}/test_data.jsonl \
--sources 'pubmedqa' \
--out-dir output \
--max-loop 3 \
--max-knowledge-loop 3 \
--max-response-loop 3 \
--gptscore-model "Alpaca_Lora" \
--demo-num 1 \
--threshold-entailment 0.8 \
--threshold-fact -1.0 \
--threshold-consistency -5

ChatGPT

navigate to the ChatGPT folder.

Directly Generate (Baseline)

CUDA_VISIBLE_DEVICES=0 python generate.py \
--input_file input.jsonl \
--out_file output.jsonl

Generate with Self-Reflection Loop (Ours)

CUDA_VISIBLE_DEVICES=0 python3 loop.py \
--input-file dataset/{source}/test_data.jsonl \
--sources 'pubmedqa' \
--out-dir output \
--max-loop 3 \
--max-knowledge-loop 3 \
--max-response-loop 3 \
--demo-num 1 \
--threshold-entailment 0.8 \
--threshold-fact -1 \
--threshold-consistency -5 

MedAlpaca

Clone the official repository and navigate to the medAlpaca folder.

Directly Generate (Baseline)

CUDA_VISIBLE_DEVICES=0 python generate.py \
--input_file input.jsonl \
--out_file output.jsonl

Generate with Self-Reflection Loop (Ours)

CUDA_VISIBLE_DEVICES=0 python3 loop.py \
--input-file dataset/{source}/test_data.jsonl \
--sources 'pubmedqa' \
--out-dir output \
--max-loop 3 \
--max-knowledge-loop 3 \
--max-response-loop 3 \
--demo-num 1 \
--threshold-entailment 0.8 \
--threshold-fact -1 \
--threshold-consistency -5

Robin-Medical

Clone the official repository and navigate to the LMFlow folder.

Directly Generate (Baseline)

CUDA_VISIBLE_DEVICES=0 python generate.py \
--input_file input.jsonl \
--out_file output.jsonl

Metrics

GPTScore

Please refers to GPTScore

MedNLI

CUDA_VISIBLE_DEVICES=0 python compute_MedNLI.py \
--data_file generated_answers.jsonl \
--out_file MedNLI_results.jsonl 

CTRLEval

git clone https://github.com/thu-coai/CTRLEval

cd CTRLEval
CUDA_VISIBLE_DEVICES=0 python compute_CTRL.py \
--data_file generated_answers.jsonl \
--out_file CTRLEval_results.csv