tensorkelechi/arxiv-cstext · Datasets at Hugging Face

text stringlengths 0 89.3k
MediTron 44 deduplicates literature samples by matching
paper titles Apollo 54 performs deduplication on exam data
by removing samples with at least 64 consecutive identical
characters to prevent data leakage and ensure accurate assess
ment results
2 Data Formatting The continued pretraining phase re
quires unstructured data the instruction finetuning phase
requires QA pairs both singleround and multiround and
the human alignment phase involves human preference data
where preferences are labeled on top of the QA pairs The
collected data may not be directly usable for the training
stages necessitating transformation into the appropriate for
mat We summarize the methods used by medical LLMs
work to convert data from various structures into three types
Unstructured Data Instruction Data and Human Preference
Data
Unstructured Data sare sourced from medical encyclo
pedias textbooks academic literature and guidelines These
data are used in the continued pretraining phase of medical
LLMs to enrich domain knowledge and enhance represen
tation capabilities After a simple truncation process these
unstructured data can effectively support the models contin
ued pretraining However structured data needs to be trans
formed into unstructured data for continued pretraining For
knowledge graph data entities relationships and attributes are
extracted Subsequently these information can be transformed
into unstructured natural language text using humandesigned
templates or natural language generation tools like LLM
AntGLMMed 46 samples subgraphs at different scales of
the knowledge graph and rewrites them as natural language
text Questioanswer pairs and NLP task data are commonly
stored in JSON format This format facilitates the extraction
of data and labels enabling the formation of unstructured
natural language text using keyvalue pairs as delimiters
MedicalGPT 35 aggregates questions and answers from the
medical encyclopedia QA dataset to create plain text data for
continued pretraining thereby integrating medical knowledgeJOURNAL OF L ATEX CLASS FILES VOL 14 NO 8 AUGUST 2021 9
Instruction Data consists of questionandanswer formats
facilitating medical LLMs in comprehending the context and
concepts of medical queries It also aids in familiarizing the
models with common expressions and language conventions
in the medical domain resulting in more natural and clini
cally relevant responses Both synthetic data and real doctor
patient conversation data adhere to the questionanswer pair
format However for sources like medical literature encyclo
pedias textbooks and knowledge graphs transformation into
questionanswer pairs requires the application of natural lan
guage generation techniques MedAlpaca 29 uses ChatGPT
to transform paragraph headings from medical textbooks into
questions and paragraph content into corresponding answers
thereby generating questionanswer pair dataset Apollo 54
partitions the corpus data into segments based on basic se
mantic units such as chapters in books paragraphs in website
content and abstracts of papers Subsequently it prompts
ChatGPT to sequentially generate questions and answers based
on these text segments InMDX 9 prompts ChatGPT to
extract fundamental medical knowledge from abstracts of med
ical literature It subsequently prompts ChatGPT to generate
corresponding questions based on these basic medical knowl
edge MeLLaMA 50 formulates specific prompts for various
datasets including medical literature clinical notes clinical
guides Wikipedia and Knowledge Graphs It then prompts
the LLM to enhance the original questions thereby creating
instruction finetuning datasets Similarly HuatuoGPTII 45
prompts ChatGPT to convert the unstructured medical corpus
into a questionanswer pair format Additionally it introduces
statistical analysis and semantic recognition techniques to
minimize discrepancies between the transformed data and the
original source Clinical Camel 33 presents a dialoguebased
knowledge encoding approach which converts dense medi
cal literature into dialogues and incorporates soft alignment
Initially it employs a teacher model to convert a segment of
literature text into a multiround dialogue with alignment con
ditions included in the prompts Subsequently a student model
is tasked with learning the mapping between the literature text
and the multiround dialogue resulting in an enhanced student
model that improves conversational capabilities BiMediX 52
utilizes samples from a multiplechoice questionandanswer
dataset and specific prompt words as inputs to ChatGPT
generating a multiround questionandanswer dataset DISC
MdeLLM 37 employs LLM to distill questions and correct
answers from the multiplechoice questionandanswer dataset
It then integrates these with explanations to produce question
andanswerformatted data
The above work employs LLMs for data formatting while
other work leverage automated methods for the same pur
pose ClinicalGPT utilizes annotations from electronic health
records as input for instruction finetuning data with diagnoses
as output forming a dataset of questionanswer pairs GPT
doctor utilizes package inserts to generate QA pairs covering
aspects such as usage dosage indications contraindications
precautions pharmacological effects chemical composition
and more The questions are constructed based on subheadings
in the package inserts like What diseases does Medicine
Name treat and then the answer is the correspondingcontent of the subheading
The medical knowledge graph embodies a wealth of struc
tured medical knowledge Efforts in medical LLMs involve
the conversion of knowledge graph data into QA pairs format
BenTsao 27 first extracts medical knowledge described in
natural language from the knowledge graph and then gen

MediTron 44 deduplicates literature samples by matching

paper titles Apollo 54 performs deduplication on exam data

by removing samples with at least 64 consecutive identical

characters to prevent data leakage and ensure accurate assess

ment results

2 Data Formatting The continued pretraining phase re

quires unstructured data the instruction finetuning phase

requires QA pairs both singleround and multiround and

the human alignment phase involves human preference data

where preferences are labeled on top of the QA pairs The

collected data may not be directly usable for the training

stages necessitating transformation into the appropriate for

mat We summarize the methods used by medical LLMs

work to convert data from various structures into three types

Unstructured Data Instruction Data and Human Preference

Data

Unstructured Data sare sourced from medical encyclo

pedias textbooks academic literature and guidelines These

data are used in the continued pretraining phase of medical

LLMs to enrich domain knowledge and enhance represen

tation capabilities After a simple truncation process these

unstructured data can effectively support the models contin

ued pretraining However structured data needs to be trans

formed into unstructured data for continued pretraining For

knowledge graph data entities relationships and attributes are

extracted Subsequently these information can be transformed

into unstructured natural language text using humandesigned

templates or natural language generation tools like LLM

AntGLMMed 46 samples subgraphs at different scales of

the knowledge graph and rewrites them as natural language

text Questioanswer pairs and NLP task data are commonly

stored in JSON format This format facilitates the extraction

of data and labels enabling the formation of unstructured

natural language text using keyvalue pairs as delimiters

MedicalGPT 35 aggregates questions and answers from the

medical encyclopedia QA dataset to create plain text data for

continued pretraining thereby integrating medical knowledgeJOURNAL OF L ATEX CLASS FILES VOL 14 NO 8 AUGUST 2021 9

Instruction Data consists of questionandanswer formats

facilitating medical LLMs in comprehending the context and

concepts of medical queries It also aids in familiarizing the

models with common expressions and language conventions

in the medical domain resulting in more natural and clini

cally relevant responses Both synthetic data and real doctor

patient conversation data adhere to the questionanswer pair

format However for sources like medical literature encyclo

pedias textbooks and knowledge graphs transformation into

questionanswer pairs requires the application of natural lan

guage generation techniques MedAlpaca 29 uses ChatGPT

to transform paragraph headings from medical textbooks into

questions and paragraph content into corresponding answers

thereby generating questionanswer pair dataset Apollo 54

partitions the corpus data into segments based on basic se

mantic units such as chapters in books paragraphs in website

content and abstracts of papers Subsequently it prompts

ChatGPT to sequentially generate questions and answers based

on these text segments InMDX 9 prompts ChatGPT to

extract fundamental medical knowledge from abstracts of med

ical literature It subsequently prompts ChatGPT to generate

corresponding questions based on these basic medical knowl

edge MeLLaMA 50 formulates specific prompts for various

datasets including medical literature clinical notes clinical

guides Wikipedia and Knowledge Graphs It then prompts

the LLM to enhance the original questions thereby creating

instruction finetuning datasets Similarly HuatuoGPTII 45

prompts ChatGPT to convert the unstructured medical corpus

into a questionanswer pair format Additionally it introduces

statistical analysis and semantic recognition techniques to

minimize discrepancies between the transformed data and the

original source Clinical Camel 33 presents a dialoguebased

knowledge encoding approach which converts dense medi

cal literature into dialogues and incorporates soft alignment

Initially it employs a teacher model to convert a segment of

literature text into a multiround dialogue with alignment con

ditions included in the prompts Subsequently a student model

is tasked with learning the mapping between the literature text

and the multiround dialogue resulting in an enhanced student

model that improves conversational capabilities BiMediX 52

utilizes samples from a multiplechoice questionandanswer

dataset and specific prompt words as inputs to ChatGPT

generating a multiround questionandanswer dataset DISC

MdeLLM 37 employs LLM to distill questions and correct

answers from the multiplechoice questionandanswer dataset

It then integrates these with explanations to produce question

andanswerformatted data

The above work employs LLMs for data formatting while

other work leverage automated methods for the same pur

pose ClinicalGPT utilizes annotations from electronic health

records as input for instruction finetuning data with diagnoses

as output forming a dataset of questionanswer pairs GPT

doctor utilizes package inserts to generate QA pairs covering

aspects such as usage dosage indications contraindications

precautions pharmacological effects chemical composition

and more The questions are constructed based on subheadings

in the package inserts like What diseases does Medicine

Name treat and then the answer is the correspondingcontent of the subheading

The medical knowledge graph embodies a wealth of struc

tured medical knowledge Efforts in medical LLMs involve

the conversion of knowledge graph data into QA pairs format

BenTsao 27 first extracts medical knowledge described in

natural language from the knowledge graph and then gen