tensorkelechi/arxiv-cstext · Datasets at Hugging Face

text stringlengths 0 89.3k
MMedLM 49 InternLM 110 7 Full Para LoRA Multilingual 022024
InMDX 9 NeuralChat 111 7 Full Para LoRA EN 022024
MeLLaMA 50 LLaMA2 101 13 70 LoRA EN 022024
JMLR 51 LLaMA2 101 7 13 EN 022024
BiMediX 52 Mixtral8x7B 112 8x7 QLoRA EN Arabic 022024
OncoGPT 53 LLaMA 99 7 LoRA EN 022024
Apollo 54Qwen 106
Gemma 113
Yi 10705 18 2
6 7 Full Para Multilingual 032024
Qibo 55 LLaMA 99 7 13 Full Para CN 032024
many PEFT methods such as Adapter Tuning 117 Prefix
Tuning 118 Cpoly 119 Prompt Tuning 120 and P
tuning 121 122 which can also reduce the performance
requirements and time consumption of the training equipment
However these methods are used less frequently than LoRA
2 Twostage Finetuning Strategy Given the variety of
data sources for instruction finetuning each representing
different aspects of medical knowledge onestage instruction
finetuning may not fully utilize these diverse sources to inject
medical instruction following capabilities into the model
Consequently some efforts in developing medical LLMs have
begun to adopt a twostage instruction finetuning approach
DISCMedLLM 37 adopted a twostage training strategy for
the instruction finetuning phase The first stage involved train
ing on largescale instruction data utilizing various datasets to
inject medical domain knowledge and dialog capabilities intothe model The second stage focused on a small amount of
highquality data generated with the assistance of ChatGPT
and filtered and guided by humans characterized by matching
human behavioral preferences The purpose of this phase
is to align the behavior of the medical LLMs with human
preferences Taiyi 42 considers that the instruction dataset
comprises multiple tasks and onestage instruction finetuning
can lead to interference between tasks necessitating a two
stage strategy The first stage uses conventional NLP task
datasets such as information extraction and text classification
The second stage uses QA and dialog task datasets training
these alongside the first stage data This twostage training al
lows the model to initially focus on processing nongenerative
tasks and then in the second phase to enhance its cross
task and dialog task processing capabilities OncoGPT 53
is first trained on Alpaca and doctorpatient communicationJOURNAL OF L ATEX CLASS FILES VOL 14 NO 8 AUGUST 2021 12
Instruction
FinetuningContinued
PretrainingHuman
Alignment
Instruction FollowingInstruction
Finetuning
Medical Dialogue
Instruction Following
Medical Dialogue
Medical Knowledge
IFTInstruction
Finetuning
Instruction Following
Medical Dialogue
CPIFT IFTHA CPIFTHA
Patient Friendly
Doctor Like
Human
AlignmentInstruction
Finetuning
Instruction Following
Medical Dialogue
Doctor Like
Patient FriendlyContinued
Pretraining
Medical Knowledge
Fig 3 Training Paradigms Details are provided on the training phases for each paradigm and the capabilities the model acquires at each phase
data from online medical consultation platforms for instruction
finetuning to gain basic medical conversational skills It is
then further trained on real doctorpatient oncology conver
sation data thoroughly proofread and reviewed by experts
and clinicians to improve cancer information processing and
conversation skills
B CPIFT
Continued pretraining also known as continued domain
adaptive pretraining involves further pretraining of LLMs on
domainspecific corpora to enrich domain knowledge 123
The CPIFT paradigm is that the general LLM initially
performs continued pretraining using unstructured medical
data to enhance its medical knowledge adapt to the specific
terminology language style and text structure of the medical
domain and gain a deeper understanding of medical expertise
and complex concepts It then performs instruction finetuning
on medical instruction data to improve its ability to follow
medical instructions There exist a few studies that follow the
CPIFT paradigm to train medical LLMs on top of general
LLMs Additionally there is a situation where some studies
44 48 49 perform continued pretraining of a general
LLM on a medical corpus However since the model does not
possess or has forgotten its instruction following capabilities at
this time instruction finetuning is conducted on the training
set of the evaluation benchmark
Theadvantage of the CPIFT paradigm is that compared to
the IFT paradigm the model learns more medical knowledge
and patterns during the continued pretraining enhancing its
generalization capabilities and enabling faster convergence
during the instruction finetuning for various medical tasks
and subdomains However the disadvantage of the CPIFT

MMedLM 49 InternLM 110 7 Full Para LoRA Multilingual 022024

InMDX 9 NeuralChat 111 7 Full Para LoRA EN 022024

MeLLaMA 50 LLaMA2 101 13 70 LoRA EN 022024

JMLR 51 LLaMA2 101 7 13 EN 022024

BiMediX 52 Mixtral8x7B 112 8x7 QLoRA EN Arabic 022024

OncoGPT 53 LLaMA 99 7 LoRA EN 022024

Apollo 54Qwen 106

Gemma 113

Yi 10705 18 2

6 7 Full Para Multilingual 032024

Qibo 55 LLaMA 99 7 13 Full Para CN 032024

many PEFT methods such as Adapter Tuning 117 Prefix

Tuning 118 Cpoly 119 Prompt Tuning 120 and P

tuning 121 122 which can also reduce the performance

requirements and time consumption of the training equipment

However these methods are used less frequently than LoRA

2 Twostage Finetuning Strategy Given the variety of

data sources for instruction finetuning each representing

different aspects of medical knowledge onestage instruction

finetuning may not fully utilize these diverse sources to inject

medical instruction following capabilities into the model

Consequently some efforts in developing medical LLMs have

begun to adopt a twostage instruction finetuning approach

DISCMedLLM 37 adopted a twostage training strategy for

the instruction finetuning phase The first stage involved train

ing on largescale instruction data utilizing various datasets to

inject medical domain knowledge and dialog capabilities intothe model The second stage focused on a small amount of

highquality data generated with the assistance of ChatGPT

and filtered and guided by humans characterized by matching

human behavioral preferences The purpose of this phase

is to align the behavior of the medical LLMs with human

preferences Taiyi 42 considers that the instruction dataset

comprises multiple tasks and onestage instruction finetuning

can lead to interference between tasks necessitating a two

stage strategy The first stage uses conventional NLP task

datasets such as information extraction and text classification

The second stage uses QA and dialog task datasets training

these alongside the first stage data This twostage training al

lows the model to initially focus on processing nongenerative

tasks and then in the second phase to enhance its cross

task and dialog task processing capabilities OncoGPT 53

is first trained on Alpaca and doctorpatient communicationJOURNAL OF L ATEX CLASS FILES VOL 14 NO 8 AUGUST 2021 12

Instruction

FinetuningContinued

PretrainingHuman

Alignment

Instruction FollowingInstruction

Finetuning

Medical Dialogue

Instruction Following

Medical Dialogue

Medical Knowledge

IFTInstruction

Finetuning

Instruction Following

Medical Dialogue

CPIFT IFTHA CPIFTHA

Patient Friendly

Doctor Like

Human

AlignmentInstruction

Finetuning

Instruction Following

Medical Dialogue

Doctor Like

Patient FriendlyContinued

Pretraining

Medical Knowledge

Fig 3 Training Paradigms Details are provided on the training phases for each paradigm and the capabilities the model acquires at each phase

data from online medical consultation platforms for instruction

finetuning to gain basic medical conversational skills It is

then further trained on real doctorpatient oncology conver

sation data thoroughly proofread and reviewed by experts

and clinicians to improve cancer information processing and

conversation skills

B CPIFT

Continued pretraining also known as continued domain

adaptive pretraining involves further pretraining of LLMs on

domainspecific corpora to enrich domain knowledge 123

The CPIFT paradigm is that the general LLM initially

performs continued pretraining using unstructured medical

data to enhance its medical knowledge adapt to the specific

terminology language style and text structure of the medical

domain and gain a deeper understanding of medical expertise

and complex concepts It then performs instruction finetuning

on medical instruction data to improve its ability to follow

medical instructions There exist a few studies that follow the

CPIFT paradigm to train medical LLMs on top of general

LLMs Additionally there is a situation where some studies

44 48 49 perform continued pretraining of a general

LLM on a medical corpus However since the model does not

possess or has forgotten its instruction following capabilities at

this time instruction finetuning is conducted on the training

set of the evaluation benchmark

Theadvantage of the CPIFT paradigm is that compared to

the IFT paradigm the model learns more medical knowledge

and patterns during the continued pretraining enhancing its

generalization capabilities and enabling faster convergence

during the instruction finetuning for various medical tasks

and subdomains However the disadvantage of the CPIFT