tensorkelechi/arxiv-cstext · Datasets at Hugging Face

text stringlengths 0 89.3k
erates the corresponding instructions and answers based on
this knowledge using ChatGPT DISCMedLLM 37 initially
prompts ChatGPT to convert diseaserelated knowledge into
theinstruction knowledge format and then prompts Chat
GPT to convert these data into a singleround medical dialogue
scenario user inquiry physician response This process
aims to augment the diversity and linguistic depth of training
datasets ClinicalGPT 36 manually designed the correspond
ing templates to convert each triple from the knowledge graph
into data in QA pair format Likewise PMCLLaMA 30
formulated QA templates to either describe an entity or predict
the relationship between two entities
Human Preference Data is intended to be used to train
the reward model and to guide the model in aligning human
preferences such as being more doctorlike patientfriendly
truthful harmless and professional This type parallels the
QA format where responses are categorized into chosen and
rejected based on their alignment with the predefined human
preferences The construction of the human preference dataset
is more timeconsuming compared to the instruction fine
tuning dataset Therefore many studies leverage ChatGPT to
facilitate the construction of the human preference dataset Hu
atuoGPT 8 utilizes real dialogues to stimulate the generative
model into producing diverse responses These responses un
dergo evaluation via ChatGPT considering coherence consis
tency with human preferences and comparison with responses
from real doctors The resultant human preference dataset is
structured based on the ratings provided by ChatGPT ChiMed
GPT 43 and MedicalGPT 35 sample questions from the
instruction finetuning dataset then the chosen response is
provided by the physician and the rejected response is pro
vided by the medical LLM BenTsao 27 To pursue high
quality human preference data some of the work uses manual
annotation Zhongjing 38 employs medical graduate students
or clinicians to annotate the human preference dataset with
nine dimensions Accuracy Safety Ethics Comprehension
Clarity Initiative Coherence Consistency and Warm Tone
yielding a dataset of superior quality Similarly ClinicalGPT
36 employs manual annotation to rank responses to medical
inquiries
3 Data Augmentation Short colloquial responses from
doctors and multiplechoice QA datasets lacking explana
tions in the collected conversation data may compromise
the accuracy and professionalism of medical LLMs Con
sequently some studies use LLMs to augment the training
dataset for addressing these issues BianQue 39 prompts
ChatGPT to refine doctors responses over multiple rounds of
conversations enhancing their empathy professionalism and
coherence DISCMedLLM 37 prompts ChatGPT to expand
on the original responses adding detail and logic without
altering their meaning It also eliminates colloquial expressions
and revises responses that do not align with the style ofJOURNAL OF L ATEX CLASS FILES VOL 14 NO 8 AUGUST 2021 10
the robot doctor such as advising patients to make appoint
ments Zhongjing 38 employs a selfinstruct 96 method to
standardize professionalize and enhance the friendliness of
doctors responses while preserving the diversity and integrity
of the original data PMCLLaMA 30 prompts GPT4 to
augment medical conversation data by generating synonymous
sentences thereby improving the models robustness to differ
ent instructions Additionally it uses ChatGPT to add rationale
explanations to multiplechoice QA data Clinical Camel 33
prompts the GPT4 to retrieve textual resources related to
the question and then supplements the multiplechoice QA
data with answer explanations according to those resources
Some studies have used LLMs to augment human preference
datasets ChiMedGPT 43 enhanced the human preference
dataset by incorporating responses from GPT4 and GPT35
The response format changed from chosen response rejected
response tochosen response GPT4s response GPT35s
response rejected response with the degree of alignment to
human preferences decreasing in that order
4 Translation To make medical LLMs widely accessible
across various countries and regions addressing language
differences is crucial Translating existing medical corpora can
overcome this barrier and maximize the utilization of high
quality data Consequently some medical LLMs work are now
focusing on ensuring the accuracy and reliability of translation
results DoctorGLM 26 prompts ChatGPT to translate the
medical corpus data to the target language and then trains
a new medical translation model for learning the mapping
from the source language to the target language BiMediX 52
employs ChatGPT for EnglishtoArabic translation followed
by a quality assessment using ChatGPT to ensure terminology
accuracy If translations do not meet a preset standard an
improvement process begins This involves providing Chat
GPT with the original English text the current translation
and its evaluation score prompting it to refine the translation
for better consistency with the original The translation quality
is continuously enhanced through iterative feedback Transla
tions that remain unsatisfactory are reviewed and corrected by
Arabicspeaking medical experts Finally a random sample of
highscoring translations is reviewed to ensure data quality To
maintain uniformity in language and style within the training
corpus HuatuoGPTII 45 was inspired by backtranslation
to standardize the format of the training corpus by prompting
ChatGPT In contrast the multilingual medical LLM Apollo
54 considers cultural differences beliefs and taboo phrases
across regions It uses local language medical datasets without
translation to avoid potential conflicts
III T RAINING PARADIGMS

erates the corresponding instructions and answers based on

this knowledge using ChatGPT DISCMedLLM 37 initially

prompts ChatGPT to convert diseaserelated knowledge into

theinstruction knowledge format and then prompts Chat

GPT to convert these data into a singleround medical dialogue

scenario user inquiry physician response This process

aims to augment the diversity and linguistic depth of training

datasets ClinicalGPT 36 manually designed the correspond

ing templates to convert each triple from the knowledge graph

into data in QA pair format Likewise PMCLLaMA 30

formulated QA templates to either describe an entity or predict

the relationship between two entities

Human Preference Data is intended to be used to train

the reward model and to guide the model in aligning human

preferences such as being more doctorlike patientfriendly

truthful harmless and professional This type parallels the

QA format where responses are categorized into chosen and

rejected based on their alignment with the predefined human

preferences The construction of the human preference dataset

is more timeconsuming compared to the instruction fine

tuning dataset Therefore many studies leverage ChatGPT to

facilitate the construction of the human preference dataset Hu

atuoGPT 8 utilizes real dialogues to stimulate the generative

model into producing diverse responses These responses un

dergo evaluation via ChatGPT considering coherence consis

tency with human preferences and comparison with responses

from real doctors The resultant human preference dataset is

structured based on the ratings provided by ChatGPT ChiMed

GPT 43 and MedicalGPT 35 sample questions from the

instruction finetuning dataset then the chosen response is

provided by the physician and the rejected response is pro

vided by the medical LLM BenTsao 27 To pursue high

quality human preference data some of the work uses manual

annotation Zhongjing 38 employs medical graduate students

or clinicians to annotate the human preference dataset with

nine dimensions Accuracy Safety Ethics Comprehension

Clarity Initiative Coherence Consistency and Warm Tone

yielding a dataset of superior quality Similarly ClinicalGPT

36 employs manual annotation to rank responses to medical

inquiries

3 Data Augmentation Short colloquial responses from

doctors and multiplechoice QA datasets lacking explana

tions in the collected conversation data may compromise

the accuracy and professionalism of medical LLMs Con

sequently some studies use LLMs to augment the training

dataset for addressing these issues BianQue 39 prompts

ChatGPT to refine doctors responses over multiple rounds of

conversations enhancing their empathy professionalism and

coherence DISCMedLLM 37 prompts ChatGPT to expand

on the original responses adding detail and logic without

altering their meaning It also eliminates colloquial expressions

and revises responses that do not align with the style ofJOURNAL OF L ATEX CLASS FILES VOL 14 NO 8 AUGUST 2021 10

the robot doctor such as advising patients to make appoint

ments Zhongjing 38 employs a selfinstruct 96 method to

standardize professionalize and enhance the friendliness of

doctors responses while preserving the diversity and integrity

of the original data PMCLLaMA 30 prompts GPT4 to

augment medical conversation data by generating synonymous

sentences thereby improving the models robustness to differ

ent instructions Additionally it uses ChatGPT to add rationale

explanations to multiplechoice QA data Clinical Camel 33

prompts the GPT4 to retrieve textual resources related to

the question and then supplements the multiplechoice QA

data with answer explanations according to those resources

Some studies have used LLMs to augment human preference

datasets ChiMedGPT 43 enhanced the human preference

dataset by incorporating responses from GPT4 and GPT35

The response format changed from chosen response rejected

response tochosen response GPT4s response GPT35s

response rejected response with the degree of alignment to

human preferences decreasing in that order

4 Translation To make medical LLMs widely accessible

across various countries and regions addressing language

differences is crucial Translating existing medical corpora can

overcome this barrier and maximize the utilization of high

quality data Consequently some medical LLMs work are now

focusing on ensuring the accuracy and reliability of translation

results DoctorGLM 26 prompts ChatGPT to translate the

medical corpus data to the target language and then trains

a new medical translation model for learning the mapping

from the source language to the target language BiMediX 52

employs ChatGPT for EnglishtoArabic translation followed

by a quality assessment using ChatGPT to ensure terminology

accuracy If translations do not meet a preset standard an

improvement process begins This involves providing Chat

GPT with the original English text the current translation

and its evaluation score prompting it to refine the translation

for better consistency with the original The translation quality

is continuously enhanced through iterative feedback Transla

tions that remain unsatisfactory are reviewed and corrected by

Arabicspeaking medical experts Finally a random sample of

highscoring translations is reviewed to ensure data quality To

maintain uniformity in language and style within the training

corpus HuatuoGPTII 45 was inspired by backtranslation

to standardize the format of the training corpus by prompting

ChatGPT In contrast the multilingual medical LLM Apollo

54 considers cultural differences beliefs and taboo phrases

across regions It uses local language medical datasets without

translation to avoid potential conflicts

III T RAINING PARADIGMS