tensorkelechi/arxiv-cstext · Datasets at Hugging Face

text stringlengths 0 89.3k
MedicalGPT 35 consist of a question a chosen response
and a rejected response The chosen response aligns with
human values more like a doctors response whereas the
rejected response is not consistent with human values com
pared to chosen response These datasets typically require
manual labeling or verification resulting in a smaller number
of public datasets of this type compared to other types Such
datasets are used in the human alignment training phase of
medical LLMs to ensure that the models responses align with
human linguistic conventions and exhibit characteristics that
are doctorlike patientfriendly harmless and professional
Knowledge Graphs represent concepts entities and their
interrelationships in the real world in a structured format
aligning information more closely with human cognition They
enhance the organization management and comprehension
of vast data Medical knowledge graphs usually contain rich
structured medical knowledge with low data noise This type
of data can motivate medical LLMs to more accurately process
medical terminology and causal relationships thus improving
models performance in reasoning medical tasks Common
medical knowledge graphs include UMLS 69 CMeKG 70
and BIOS 71 Since these are not in QA pairs or plain text
format a carefully designed data transformation pipeline is
required to convert them
NLP Task datasets include Named Entity Recognition
NER Relation Extraction RE Causal Relation Extraction
CRE Text Classification TC Text PairSemantic Similarity
TPSS Text PairTextual Entailment TPTE and others
These datasets enhance model performance in various medical
tasks and instruction following MeLLaMA 50 and Taiyi
42 utilize these conventional medical NLP datasets to im
prove their models generalization capabilities It is important
to note that these data generally cannot be used directly for
training medical LLMs and must be transformed into plain
text or questionanswer pair format by data processing
Public datasets are easily accessible and high quality but
they do not all meet the needs of training medical LLMs inthe medical domain such as focusing on pediatrics or other
nuanced domains of medical LLMs It is necessary to create
an individual training set to meet customized needs
2 Public Medical Corpus Publicly accessible online med
ical data is a primary source for training medical LLMs This
data encompasses diverse medical information from various
regions populations and platforms While it offers extensive
coverage and a wide range of medical knowledge it often
presents challenges related to privacy and data quality necessi
tating further cleaning and processing Here we summarize six
common data sources for medical LLMs medical encyclope
dias medical textbooks medical academic literature medical
websites medical guidelines and package inserts
Medical Encyclopedias are comprehensive reference re
sources typically available online that systematically docu
ment a wide range of medical knowledge including diseases
symptoms diagnostic methods treatments and related med
ical technologies Compiled by professional medical groups
academic institutions or authoritative publishers their content
undergoes rigorous review and regular updates to ensure
accuracy and currency The structured organization of these
encyclopedias facilitates efficient data retrieval and processing
Notable medical encyclopedia websites include MedlinePlus
72 Mayo Clinic 73 WebMD 74 Medical LLMs often
use encyclopedias as a training data source to enhance their
medical knowledge 45 54 Although the format of ency
clopedia content is suitable for the pretraining of these models
further data processing and transformation are necessary for
instruction finetuning
Medical Textbooks provide a systematic introduction to
medical knowledge and skills serving as comprehensive ref
erences for medical students physicians nurses and other
healthcare professionals They cover a wide range of topics
from basic sciences eg anatomy physiology biochem
istry to clinical practice eg internal medicine surgery
pediatrics obstetrics and gynecology Authored by experts
and peerreviewed these textbooks ensure the accuracy and
authority of medical information Common sources include
FreeBooks4Doctors 75 Internet Archive 76 and Bookboon
77 Similar to medical encyclopedias medical textbooks of
fer extensive specialized and structured information Conse
quently medical LLMs 46 49 51 55 utilize these text
books as training data to incorporate professional and precise
medical knowledge However the specialized terminology and
complex linguistic structures in medical textbooks can limit
their accessibility for users without a medical background
Medical Academic Literature encompasses specialized
publications that research explore and discuss various fields
of medicine This literature includes peerreviewed journal
articles scholarly books conference papers and research re
ports Covering specialties such as clinical medicine biomed
ical sciences public health pharmacology and physiology
these publications provide uptodate medical knowledge and
evidence for researchers clinicians students and allied profes
sionals Common sources include PubMed 78 Embase 79
and ScienceDirect 80 Due to its specialization accuracy
and currency medical academic literature is frequently used to
train medical LLMs 30 44 50 enhancing their medicalJOURNAL OF L ATEX CLASS FILES VOL 14 NO 8 AUGUST 2021 6
expertise However the specialized terminology and technical
content can make these models less accessible to the general
public Additionally acquiring academic literature data can
present versioning challenges
Medical Websites encompass various online platforms
such as forums consultations news and educational re

MedicalGPT 35 consist of a question a chosen response

and a rejected response The chosen response aligns with

human values more like a doctors response whereas the

rejected response is not consistent with human values com

pared to chosen response These datasets typically require

manual labeling or verification resulting in a smaller number

of public datasets of this type compared to other types Such

datasets are used in the human alignment training phase of

medical LLMs to ensure that the models responses align with

human linguistic conventions and exhibit characteristics that

are doctorlike patientfriendly harmless and professional

Knowledge Graphs represent concepts entities and their

interrelationships in the real world in a structured format

aligning information more closely with human cognition They

enhance the organization management and comprehension

of vast data Medical knowledge graphs usually contain rich

structured medical knowledge with low data noise This type

of data can motivate medical LLMs to more accurately process

medical terminology and causal relationships thus improving

models performance in reasoning medical tasks Common

medical knowledge graphs include UMLS 69 CMeKG 70

and BIOS 71 Since these are not in QA pairs or plain text

format a carefully designed data transformation pipeline is

required to convert them

NLP Task datasets include Named Entity Recognition

NER Relation Extraction RE Causal Relation Extraction

CRE Text Classification TC Text PairSemantic Similarity

TPSS Text PairTextual Entailment TPTE and others

These datasets enhance model performance in various medical

tasks and instruction following MeLLaMA 50 and Taiyi

42 utilize these conventional medical NLP datasets to im

prove their models generalization capabilities It is important

to note that these data generally cannot be used directly for

training medical LLMs and must be transformed into plain

text or questionanswer pair format by data processing

Public datasets are easily accessible and high quality but

they do not all meet the needs of training medical LLMs inthe medical domain such as focusing on pediatrics or other

nuanced domains of medical LLMs It is necessary to create

an individual training set to meet customized needs

2 Public Medical Corpus Publicly accessible online med

ical data is a primary source for training medical LLMs This

data encompasses diverse medical information from various

regions populations and platforms While it offers extensive

coverage and a wide range of medical knowledge it often

presents challenges related to privacy and data quality necessi

tating further cleaning and processing Here we summarize six

common data sources for medical LLMs medical encyclope

dias medical textbooks medical academic literature medical

websites medical guidelines and package inserts

Medical Encyclopedias are comprehensive reference re

sources typically available online that systematically docu

ment a wide range of medical knowledge including diseases

symptoms diagnostic methods treatments and related med

ical technologies Compiled by professional medical groups

academic institutions or authoritative publishers their content

undergoes rigorous review and regular updates to ensure

accuracy and currency The structured organization of these

encyclopedias facilitates efficient data retrieval and processing

Notable medical encyclopedia websites include MedlinePlus

72 Mayo Clinic 73 WebMD 74 Medical LLMs often

use encyclopedias as a training data source to enhance their

medical knowledge 45 54 Although the format of ency

clopedia content is suitable for the pretraining of these models

further data processing and transformation are necessary for

instruction finetuning

Medical Textbooks provide a systematic introduction to

medical knowledge and skills serving as comprehensive ref

erences for medical students physicians nurses and other

healthcare professionals They cover a wide range of topics

from basic sciences eg anatomy physiology biochem

istry to clinical practice eg internal medicine surgery

pediatrics obstetrics and gynecology Authored by experts

and peerreviewed these textbooks ensure the accuracy and

authority of medical information Common sources include

FreeBooks4Doctors 75 Internet Archive 76 and Bookboon

77 Similar to medical encyclopedias medical textbooks of

fer extensive specialized and structured information Conse

quently medical LLMs 46 49 51 55 utilize these text

books as training data to incorporate professional and precise

medical knowledge However the specialized terminology and

complex linguistic structures in medical textbooks can limit

their accessibility for users without a medical background

Medical Academic Literature encompasses specialized

publications that research explore and discuss various fields

of medicine This literature includes peerreviewed journal

articles scholarly books conference papers and research re

ports Covering specialties such as clinical medicine biomed

ical sciences public health pharmacology and physiology

these publications provide uptodate medical knowledge and

evidence for researchers clinicians students and allied profes

sionals Common sources include PubMed 78 Embase 79

and ScienceDirect 80 Due to its specialization accuracy

and currency medical academic literature is frequently used to

train medical LLMs 30 44 50 enhancing their medicalJOURNAL OF L ATEX CLASS FILES VOL 14 NO 8 AUGUST 2021 6

expertise However the specialized terminology and technical

content can make these models less accessible to the general

public Additionally acquiring academic literature data can

present versioning challenges

Medical Websites encompass various online platforms

such as forums consultations news and educational re