tensorkelechi/arxiv-cstext · Datasets at Hugging Face

text stringlengths 0 89.3k
Transitioning from general LLMs to a medical LLMs in
volves three stages of continued training continued pretrain
ing instruction finetuning and human alignment However
given that the general LLM has completed training and pos
sesses some medical knowledge alongside its general knowl
edge all three phases may not be necessary to train a medical
LLM based on the general LLM Tab II presents information
about the medical LLMs including number of parametersbase model and training stages employed Medical LLMs
learn different capabilities at various training stages and
have diverse demands on computational resources and dataset
scales We categorize the training process into four paradigms
based on the combination of training stages as depicted in
Fig 3 The characteristics and strategies of each paradigm are
described in detail below
A IFT
Instruction finetuning also termed supervised finetuning
refers to parameter finetuning of pretrained LLMs using
instruction data QA pairs During the training process only
the loss of response content is considered After instruction
finetuning medical LLMs can demonstrate strong medical
instruction following and zeroshot learning ability It can
effectively grasp instructions from medical practitioners or
patients thereby enhancing doctors clinical efficiency and
addressing patients needs efficiently
Medical LLMs depicted in Tab II all employ instruction
finetuning to enhance their capacity for addressing medical
issues This is because the general LLMs inherently possess
extensive knowledge and thus models only require instruc
tion finetuning to adapt to medical directives Additionally
the emergence of Parameter Efficient FineTuning PEFT
accelerates the training process of instruction finetuning and
reduces the demand for computational resources Strengths
Instruction finetuning IFT can improve large medical lan
guage models performance in instruction following adherence
with minimal computational resources and time consumption
Weaknesses Without continued pretraining and human align
ment the knowledge understanding of the model may be
limited and the response of the model may not accurately
simulate the characteristics of a physician
1 Parameter Efficient FineTuning As can be seen from
Tab II there are two types of instruction finetuning methods
used in medical LLMs one is full parameter finetuning and
the other is Parameter Efficient FineTuning PEFT 114
Full parameter finetuning involves training and updating all
parameters related to the medical LLM during instruction fine
tuning Conversely PEFT reduces the number of model pa
rameters requiring training thereby reducing the performance
demands of the necessary equipment for training while en
suring that the finetuned models performance is comparable
to that of full parameter finetuning The most commonly
used parameterefficient finetuning method in the medical
LLMs is LowRank Adaptation LoRA 115 LoRA approx
imates parameter updates for each layer by adding lowrank
decomposition matrices while freezing the original weights
thereby reducing the number of training parameters and the
training time Another commonly used method is QLoRA
116 which is a parameter quantization technique based on
the LoRA method In QLoRA the original model parameters
are quantized to 4bit while the newly added lowrank de
composition matrix parameters are saved as 16bit for training
This approach further reduces the performance requirements of
the training equipment and lowers the threshold for developing
medical LLMs in medical institutions In addition there areJOURNAL OF L ATEX CLASS FILES VOL 14 NO 8 AUGUST 2021 11
TABLE II
DETAILED INFORMATION OF MEDICAL LARGE LANGUAGE MODELS NOTE P ARADENOTES PARAMETERS CP DENOTES CONTINUED
PRETRAINING IFT DENOTES INSTRUCTION FINE TUNING AND HA DENOTES HUMAN ALIGNMENT
Models Backbone Para B CP IFT IFT Methods HAPreferred
LanguagesOpen
SourcesData
MedPaLM 24 PaLM 98 540 Prompt Tuning EN 012023
ChatDoctor 25 LLaMA 99 7 Full Para EN 032023
DoctorGLM 26 ChatGLM 100 6 LoRA CN 042023
BenTsao 27 LLaMA 99 7 LoRA CN 042023
ChatGLMMed 28 ChatGLM 100 6 LoRA CN 042023
MedAlpaca 29 LLaMA 99 7 13 Full Para LoRA CN 042023
PMCLLaMA 30 LLaMA2 101 13 Full Para CN 042023
HuatuoGPT 8Baichuan 102
ZiyaLLaMA 1037 13 Full Para CN 052023
ChatMedConsult
31LLaMA 99 7 LoRA CN 052023
MedPaLM 2 32 PaLM2 104 EN 052023
Clinical Camel 33 LLaMA2 101 13 70 QLoRA EN 052023
ShenNongTCM 34 LLaMA 99 7 LoRA CN 062023
MedicalGPT 35ZiyaLLaMA 103
BaichuanChat 10213 LoRA EN CN 062023
ClinicalGPT 36 BLOOM 105 7 LoRA CN 062023
DISCMedLLM 37 Baichuan 102 13 Full Para CN 082023
Zhongjing 38 ZiyaLLaMA 103 13 LoRA CN 082023
BianQue 39 ChatGLM 100 6 Full Para CN 102023
Alpacare 40 LLaMA 99 7 13 Full Para EN 102023
QilinMed 41 Baichuan 102 7 LoRA CN 102023
Taiyi 42 Qwen 106 7 QLoRA ENCN 112023
ChiMedGPT 43 ZiyaLLaMA 103 13 Full Para CN 112023
MediTron 44 LLaMA2 101 7 70 Full Para EN 112023
HuatuoGPTII 45Baichuan2 102
Yi 1077 13 34 Full Para CN 122023
AntGLMMed 46 GLM 108 10 Full Para LoRA
CpolyEN CN 122023
GPTdoctor 47 Baichuan2Chat 102 13 LoRA CN 122023
BioMistral 48 MistralInstruct 109 7 QLoRA Multilingual 022024

Transitioning from general LLMs to a medical LLMs in

volves three stages of continued training continued pretrain

ing instruction finetuning and human alignment However

given that the general LLM has completed training and pos

sesses some medical knowledge alongside its general knowl

edge all three phases may not be necessary to train a medical

LLM based on the general LLM Tab II presents information

about the medical LLMs including number of parametersbase model and training stages employed Medical LLMs

learn different capabilities at various training stages and

have diverse demands on computational resources and dataset

scales We categorize the training process into four paradigms

based on the combination of training stages as depicted in

Fig 3 The characteristics and strategies of each paradigm are

described in detail below

A IFT

Instruction finetuning also termed supervised finetuning

refers to parameter finetuning of pretrained LLMs using

instruction data QA pairs During the training process only

the loss of response content is considered After instruction

finetuning medical LLMs can demonstrate strong medical

instruction following and zeroshot learning ability It can

effectively grasp instructions from medical practitioners or

patients thereby enhancing doctors clinical efficiency and

addressing patients needs efficiently

Medical LLMs depicted in Tab II all employ instruction

finetuning to enhance their capacity for addressing medical

issues This is because the general LLMs inherently possess

extensive knowledge and thus models only require instruc

tion finetuning to adapt to medical directives Additionally

the emergence of Parameter Efficient FineTuning PEFT

accelerates the training process of instruction finetuning and

reduces the demand for computational resources Strengths

Instruction finetuning IFT can improve large medical lan

guage models performance in instruction following adherence

with minimal computational resources and time consumption

Weaknesses Without continued pretraining and human align

ment the knowledge understanding of the model may be

limited and the response of the model may not accurately

simulate the characteristics of a physician

1 Parameter Efficient FineTuning As can be seen from

Tab II there are two types of instruction finetuning methods

used in medical LLMs one is full parameter finetuning and

the other is Parameter Efficient FineTuning PEFT 114

Full parameter finetuning involves training and updating all

parameters related to the medical LLM during instruction fine

tuning Conversely PEFT reduces the number of model pa

rameters requiring training thereby reducing the performance

demands of the necessary equipment for training while en

suring that the finetuned models performance is comparable

to that of full parameter finetuning The most commonly

used parameterefficient finetuning method in the medical

LLMs is LowRank Adaptation LoRA 115 LoRA approx

imates parameter updates for each layer by adding lowrank

decomposition matrices while freezing the original weights

thereby reducing the number of training parameters and the

training time Another commonly used method is QLoRA

116 which is a parameter quantization technique based on

the LoRA method In QLoRA the original model parameters

are quantized to 4bit while the newly added lowrank de

composition matrix parameters are saved as 16bit for training

This approach further reduces the performance requirements of

the training equipment and lowers the threshold for developing

medical LLMs in medical institutions In addition there areJOURNAL OF L ATEX CLASS FILES VOL 14 NO 8 AUGUST 2021 11

TABLE II

DETAILED INFORMATION OF MEDICAL LARGE LANGUAGE MODELS NOTE P ARADENOTES PARAMETERS CP DENOTES CONTINUED

PRETRAINING IFT DENOTES INSTRUCTION FINE TUNING AND HA DENOTES HUMAN ALIGNMENT

Models Backbone Para B CP IFT IFT Methods HAPreferred

LanguagesOpen

SourcesData

MedPaLM 24 PaLM 98 540 Prompt Tuning EN 012023

ChatDoctor 25 LLaMA 99 7 Full Para EN 032023

DoctorGLM 26 ChatGLM 100 6 LoRA CN 042023

BenTsao 27 LLaMA 99 7 LoRA CN 042023

ChatGLMMed 28 ChatGLM 100 6 LoRA CN 042023

MedAlpaca 29 LLaMA 99 7 13 Full Para LoRA CN 042023

PMCLLaMA 30 LLaMA2 101 13 Full Para CN 042023

HuatuoGPT 8Baichuan 102

ZiyaLLaMA 1037 13 Full Para CN 052023

ChatMedConsult

31LLaMA 99 7 LoRA CN 052023

MedPaLM 2 32 PaLM2 104 EN 052023

Clinical Camel 33 LLaMA2 101 13 70 QLoRA EN 052023

ShenNongTCM 34 LLaMA 99 7 LoRA CN 062023

MedicalGPT 35ZiyaLLaMA 103

BaichuanChat 10213 LoRA EN CN 062023

ClinicalGPT 36 BLOOM 105 7 LoRA CN 062023

DISCMedLLM 37 Baichuan 102 13 Full Para CN 082023

Zhongjing 38 ZiyaLLaMA 103 13 LoRA CN 082023

BianQue 39 ChatGLM 100 6 Full Para CN 102023

Alpacare 40 LLaMA 99 7 13 Full Para EN 102023

QilinMed 41 Baichuan 102 7 LoRA CN 102023

Taiyi 42 Qwen 106 7 QLoRA ENCN 112023

ChiMedGPT 43 ZiyaLLaMA 103 13 Full Para CN 112023

MediTron 44 LLaMA2 101 7 70 Full Para EN 112023

HuatuoGPTII 45Baichuan2 102

Yi 1077 13 34 Full Para CN 122023

AntGLMMed 46 GLM 108 10 Full Para LoRA

CpolyEN CN 122023

GPTdoctor 47 Baichuan2Chat 102 13 LoRA CN 122023

BioMistral 48 MistralInstruct 109 7 QLoRA Multilingual 022024