text
stringlengths 0
89.3k
|
---|
data cleaning such as filtering for medicalrelated content The
|
cleaned data can then be used to continued pretraining Medical
|
LLMs enhancing their understanding of the distributional
|
and statistical characteristics of medical data and thereby
|
improving the stability and robustness of the models
|
QuestionAnswer Pairs QA include singleturn question
|
answer multiturn questionanswer conversations and
|
multiplechoice questionanswer datasets Singleturn
|
questionanswer datasets such as Huatuo26M 62 and
|
ChiMed 63 primarily consist of questions and answers
|
collected from reallife scenarios Each question addresses aJOURNAL OF L ATEX CLASS FILES VOL 14 NO 8 AUGUST 2021 4
|
TABLE I
|
CORPUS SOURCES PROCESSING METHODS TRAINING SETFORMAT AND SCALE FOR MEDICAL LLM S THE ABBREVIATIONS HERE ARE PUB DATA
|
EXISTING PUBLIC DATASETS PUB CORPUS PUBLIC MEDICAL CORPUS ORG CORPUS PROFESSIONAL MEDICAL ORGANIZATION CORPUS
|
SYNSYNTHETIC DATACLEAN D ATA CLEANING FORMAT D ATA FORMATTING C ONSTRUCTION AUG D ATA AUGMENTATION
|
TRANS TRANSLATION QA Q UESTION ANSWER DATASETS SINGLE TURN OR MULTIPLE TURNS OR MULTIPLE CHOICE UD U NSTRUCTURED DATA
|
HPD H UMAN PREFERENCE DATA KG K NOWLEDGE GRAPH NLP NLP T ASK MB M EDICAL WEBSITE ENCYC E NCYCLOPEDIA TB T EXTBOOK
|
GUID G UIDELINE PI P ACKAGE INSERT MAC M EDICAL ACADEMIC LITERATURE CN C LINICAL NOTE EHR E LECTRONIC HEALTH RECORD MP
|
MEDICAL PRESCRIPTION DPC D OCTOR PATIENT CONVERSATION
|
ModelsTraining Corpus Sources Data Processing Standardized Training Sets
|
Pub Data Pub Corpus Org Corpus Syn Clean Format Aug Trans Format Scale
|
MedPaLM 24 QA MB InsD 65 items
|
ChatDoctor 25 MB InsD 100K items
|
DoctorGLM 26 InsD 4487K items
|
BenTsao 27 KG InsD 8K items
|
ChatGLMMed 28 KG InsD 76K items
|
MedAlpaca 29 QA NLP MB InsD 248K items
|
PMCLLaMA 30 UD QA KG MAL TB UD 79B tokens
|
InsD 202M tokens
|
HuatuoGPT 8 QA DPC InsD 226K items
|
ChatMedConsult 31 MB InsD 110K items
|
MedPaLM 2 32 QA InsD 193K items
|
Clinical Camel 33 QA MAL InsD 174K items
|
ShenNongTCM 34 KG InsD 113K items
|
MedicalGPT 35 QA HPDUD 361K items
|
InsD 207M items
|
HPD 38K items
|
ClinicalGPT 36 QA KG EHRInsD 423K items
|
HPD 10K items
|
DISCMedLLM 37 QA KG InsD 514K items
|
Zhongjing 38QA KG
|
NLPTBEHR CN
|
DPC UD 1086 MB
|
InsD 470K items
|
HPD 20K items
|
BianQue 39 DPC InsD 2437K items
|
Alpacare 40 InsD 52K items
|
QilinMed 41QA KG
|
HPDUD 2B tokens
|
InsD 11M items
|
HPD 6K items
|
Taiyi 42 QA NLP InsD 1114K items
|
ChiMedGPT 43UD QA
|
HPD UD 369K items
|
InsD 1269K items
|
HPD 4K items
|
MediTron 44 UD QA MAL CN UD 48 tokens
|
InsD 369K items
|
HuatuoGPTII 45 UDEncyc TB
|
MAL MW InsD 5394K items
|
AntGLMMed 46 UD QA KG TB MAL DPC UD 1539B tokens
|
InsD 632K items
|
GPTdoctor 47 MB PI InsD 1939K items
|
BioMistral 48 UD QAUD 3B tokens
|
InsD 405K items
|
MMedLM 49 UD MW TB UD 255B tokens
|
InMDX 9 MAL UD 150M tokens
|
InsD 1701K items
|
MeLLaMA 50UD QA
|
KG NLPMAL EHR CN UD 129B tokens
|
InsD 214K items
|
JMLR 51 QA TB Guid EHR UD InsD
|
BiMediX 52 QA InsD 1311K items
|
OncoGPT 53 MB DPC InsD 332K items
|
Apollo 54 UD QATB MAL
|
MW Encyc
|
Guid UD 2054M tokens
|
InsD 481M items
|
Qibo 55 QA NLP TB Encyc MP UD 1277MB InsDJOURNAL OF L ATEX CLASS FILES VOL 14 NO 8 AUGUST 2021 5
|
specific topic within the medical domain enabling medical
|
LLMs to learn a specialized manner and style of responding
|
However due to its singleturn QA nature it falls short
|
of enhancing the ability of medical LLMs to process
|
contextual information Conversely datasets in a multiturn
|
QA format such as MedDialog 64 BianqueCorpus 39 and
|
CMtMedQA 38 have the potential to augment the capacity
|
of medical LLMs to handle contextual information and
|
simulate authentic doctorpatient interactions Additionally
|
these multiround QA datasets often include instances where
|
doctors inquire about patients conditions thereby bolstering
|
the medical LLMs ability to ask proactive questions and
|
guide users in describing their situations Multiplechoice
|
QA datasets such as MMLU 65 MedQA 66 PubMedQA
|
67 and MedMCQA 68 consist of questions and options
|
with some derived from actual medical exam questions In
|
addition to correct answers some of these datasets include
|
detailed explanations This enhances the models ability to
|
select correct answers and improves its understanding and
|
reasoning about complex medical information
|
Human Preference Data such as Zhongjing rlhf 38 and
|
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.