tensorkelechi/arxiv-cstext · Datasets at Hugging Face

text stringlengths 0 89.3k
ers medical LLMs to generate outputs that are more aligned
with medical principles and human expectations thereby
reducing potentially harmful outputs and fostering greater
caution when answering medical questions In particular the
paradigm can better protect patient privacy when dealing
with sensitive medical information Compared to the IFTHA
paradigm this CPIFTHA paradigm incorporates additional
stages of continued pretraining resulting in medical LLMs
equipped with a more robust foundation in medical knowledge
This enhanced knowledge enables models to posses greater
adaptability within the medical domain Consequently models
using this paradigm are able to provide more accurate and
insightful answers which is especially important in medical
domains that involve highly specialized and precise questions
E Summary
The choice of paradigm should be based on the demands re
sources and time constraints of a specific project In the medi
cal field where high expertise and accuracy are paramount theparadigm CPIFTHA is more advantageous However in sce
narios requiring rapid deployment with limited resources the
IFT paradigm can better meet practical needs The CPIFT and
IFTHA paradigms strike a balance between the performance
of medical LLMs and the complexity of development Specifi
cally the CPIFT paradigm focuses on the comprehension and
processing of medical knowledge ensuring that the model can
accurately grasp and apply relevant information While the
IFTHA paradigm emphasizes the safety of responses and a
doctorlike professionalism ensuring patient safety and trust
IV E VALUATIONS
Given the specialized and sensitive nature of medical in
formation erroneous or unsafe responses could potentially
mislead patients leading to harm Hence it is crucial to
accurately and comprehensively assess the accuracy of medical
LLMs We systematically summarize the existing evaluation
methods for medical LLMs and classify them into two main
categories Machine Evaluation and HumanCentric Evalua
tion The former focus on the use of deterministic evaluation
metrics to measure the performance of medical LLMs while
the latter evaluates them from a human perspective consider
ing subjective criteria Tab III presents the current evaluation
protocols and metrics for medical LLMs
A Machine Evaluation
1 Benchmarks Machine Evaluation assesses the perfor
mance of medical LLMs in handling natural language tasks
within the medical domain utilizing established benchmarks
for natural language understanding and generation These
benchmarks provide explicit static evaluation criteria that
quantitatively measure the performance of medical LLMs
As shown in Tab III few studies have used natural language
understanding tasks as evaluation benchmarks such as Medi
cal Named Entity Recognition Medical Relationship Extrac
tion and Medical Text Categorization General LLMs typically
include natural language understanding tasks in training sets
providing them with capability in this area This evaluation
aims to determine whether Medical LLMs after adaptation
to medical information retain natural language understanding
ability in the medical domain MeLLaMA 50 and Taiyi 42
utilize natural language understanding benchmarks to evaluate
the performance of the proposed medical LLM They first
perform taskspecific supervised finetuning on the training
sets of these benchmarks followed by performance evaluation
on the test sets Common evaluation metrics for these tasks
include Accuracy F1Score 128 and BERTScore 129
Many studies have used QA tasks singleturn multiturn
multiplechoice in natural language generation as evaluation
benchmarks to measure the quality of response content and the
performance processing medical QA tasks for medical LLMs
In particular the singleturn QA benchmark involves a single
question and a single answer without context or continuous
dialog It allows direct assessment of the models performance
and knowledge coverage on specific medical questions And
it is applicable to a wide range of medical topics from simple
definitions to complex diagnostic problems The multiturnJOURNAL OF L ATEX CLASS FILES VOL 14 NO 8 AUGUST 2021 14
TABLE III
EVALUATION SETTING DETAILS FOR THE MEDICAL LARGE LANGUAGE MODEL THE ABBREVIATIONS ARE AS FOLLOWS STQA FOR SINGLE TURN
QA MTQA FOR MULTIPLE TURN QA MCQA FOR MULTIPLE CHOICE QA NLP FOR CONVENTIONAL NATURAL LANGUAGE PROCESSING TASKS
ZEROFOR ZEROSHOT LEARNING FEWFOR FEWSHOT LEARNING AND SFT FOR TASKSPECIFIC SUPERVISED FINETUNING
Models TypesMachine Evaluation HumanCentric Evaluation
Metrics Protocols Evaluator Dimensions Protocols
MedPaLM 24 STQA MCQA Accuracy Zero Few HumanProfessional Safe
HelpfulIndividual
ChatDoctor 25 STQA BERTScore Zero Human Case Study
DoctorGLM 26 Human Case Study
BenTsao 27 STQA Human Safe Fluent Helpful Individual
ChatGLMMed 28 Human Case Study
MedAlpaca 29 MCQA Accuracy Zero
PMCLLaMA 30 MCQA Accuracy Zero SFT
HuatuoGPT 8 STQA MTQABLEU ROUGE
GLEU DistinctZeroHuman
LLMProfessional
Proactive Doctor
like PatientfriendlyPairwise
ChatMedConsult
31Human Case Study
MedPaLM 2 32 STQA MCQA Accuracy Zero Few HumanProfessional Safe
HelpfulIndividual
Pairwise
Clinical Camel 33 MCQA Accuracy Zero Few
ShenNongTCM 34 Human Case Study
MedicalGPT 35 Human Case Study
ClinicalGPT 36 STQA MCQABLEU ROUGE

ers medical LLMs to generate outputs that are more aligned

with medical principles and human expectations thereby

reducing potentially harmful outputs and fostering greater

caution when answering medical questions In particular the

paradigm can better protect patient privacy when dealing

with sensitive medical information Compared to the IFTHA

paradigm this CPIFTHA paradigm incorporates additional

stages of continued pretraining resulting in medical LLMs

equipped with a more robust foundation in medical knowledge

This enhanced knowledge enables models to posses greater

adaptability within the medical domain Consequently models

using this paradigm are able to provide more accurate and

insightful answers which is especially important in medical

domains that involve highly specialized and precise questions

E Summary

The choice of paradigm should be based on the demands re

sources and time constraints of a specific project In the medi

cal field where high expertise and accuracy are paramount theparadigm CPIFTHA is more advantageous However in sce

narios requiring rapid deployment with limited resources the

IFT paradigm can better meet practical needs The CPIFT and

IFTHA paradigms strike a balance between the performance

of medical LLMs and the complexity of development Specifi

cally the CPIFT paradigm focuses on the comprehension and

processing of medical knowledge ensuring that the model can

accurately grasp and apply relevant information While the

IFTHA paradigm emphasizes the safety of responses and a

doctorlike professionalism ensuring patient safety and trust

IV E VALUATIONS

Given the specialized and sensitive nature of medical in

formation erroneous or unsafe responses could potentially

mislead patients leading to harm Hence it is crucial to

accurately and comprehensively assess the accuracy of medical

LLMs We systematically summarize the existing evaluation

methods for medical LLMs and classify them into two main

categories Machine Evaluation and HumanCentric Evalua

tion The former focus on the use of deterministic evaluation

metrics to measure the performance of medical LLMs while

the latter evaluates them from a human perspective consider

ing subjective criteria Tab III presents the current evaluation

protocols and metrics for medical LLMs

A Machine Evaluation

1 Benchmarks Machine Evaluation assesses the perfor

mance of medical LLMs in handling natural language tasks

within the medical domain utilizing established benchmarks

for natural language understanding and generation These

benchmarks provide explicit static evaluation criteria that

quantitatively measure the performance of medical LLMs

As shown in Tab III few studies have used natural language

understanding tasks as evaluation benchmarks such as Medi

cal Named Entity Recognition Medical Relationship Extrac

tion and Medical Text Categorization General LLMs typically

include natural language understanding tasks in training sets

providing them with capability in this area This evaluation

aims to determine whether Medical LLMs after adaptation

to medical information retain natural language understanding

ability in the medical domain MeLLaMA 50 and Taiyi 42

utilize natural language understanding benchmarks to evaluate

the performance of the proposed medical LLM They first

perform taskspecific supervised finetuning on the training

sets of these benchmarks followed by performance evaluation

on the test sets Common evaluation metrics for these tasks

include Accuracy F1Score 128 and BERTScore 129

Many studies have used QA tasks singleturn multiturn

multiplechoice in natural language generation as evaluation

benchmarks to measure the quality of response content and the

performance processing medical QA tasks for medical LLMs

In particular the singleturn QA benchmark involves a single

question and a single answer without context or continuous

dialog It allows direct assessment of the models performance

and knowledge coverage on specific medical questions And

it is applicable to a wide range of medical topics from simple

definitions to complex diagnostic problems The multiturnJOURNAL OF L ATEX CLASS FILES VOL 14 NO 8 AUGUST 2021 14

TABLE III

EVALUATION SETTING DETAILS FOR THE MEDICAL LARGE LANGUAGE MODEL THE ABBREVIATIONS ARE AS FOLLOWS STQA FOR SINGLE TURN

QA MTQA FOR MULTIPLE TURN QA MCQA FOR MULTIPLE CHOICE QA NLP FOR CONVENTIONAL NATURAL LANGUAGE PROCESSING TASKS

ZEROFOR ZEROSHOT LEARNING FEWFOR FEWSHOT LEARNING AND SFT FOR TASKSPECIFIC SUPERVISED FINETUNING

Models TypesMachine Evaluation HumanCentric Evaluation

Metrics Protocols Evaluator Dimensions Protocols

MedPaLM 24 STQA MCQA Accuracy Zero Few HumanProfessional Safe

HelpfulIndividual

ChatDoctor 25 STQA BERTScore Zero Human Case Study

DoctorGLM 26 Human Case Study

BenTsao 27 STQA Human Safe Fluent Helpful Individual

ChatGLMMed 28 Human Case Study

MedAlpaca 29 MCQA Accuracy Zero

PMCLLaMA 30 MCQA Accuracy Zero SFT

HuatuoGPT 8 STQA MTQABLEU ROUGE

GLEU DistinctZeroHuman

LLMProfessional

Proactive Doctor

like PatientfriendlyPairwise

ChatMedConsult

31Human Case Study

MedPaLM 2 32 STQA MCQA Accuracy Zero Few HumanProfessional Safe

HelpfulIndividual

Pairwise

Clinical Camel 33 MCQA Accuracy Zero Few

ShenNongTCM 34 Human Case Study

MedicalGPT 35 Human Case Study

ClinicalGPT 36 STQA MCQABLEU ROUGE