text
stringlengths 0
89.3k
|
---|
Medical Large
|
Language Model
|
Evaluations Sec IV
|
Machine Evaluation
|
Human Centric Evaluation
|
Medical
|
Datasets
|
Public
|
Corpus
|
Synthetic
|
Data
|
Organization
|
Corpus
|
Data
|
Cleaning
|
Data
|
Augmentation
|
Data
|
Formatting
|
Translation
|
Standardized
|
Training Sets
|
IFT
|
CPIFT
|
IFTHA
|
CPIFTHA
|
Accuracy Matching Degree
|
Rating
|
Win tie and loss rate
|
Fig 1 Training Pipeline from General LLMs to Medical LLMs Firstly the medical corpus is collected and processed to form a standardized training set
|
Next an appropriate training paradigm is selected to train General LLMs to become Medical LLMs with medical knowledge The training paradigms consist
|
of three optional training phases Continued Pretraining CP Instruction Finetuning IFT and Human Alignment HA Finally Medical LLMs are evaluated
|
from both machine and human perspectives
|
effectiveness and efficiency of training domain LLMs from
|
general LLMs this paper systematically surveys medical
|
LLMs derived from further training of general LLMs from
|
the perspectives of dataset methodology and evaluation
|
The main content is depicted in Fig 1 Firstly we focus
|
on the construction of training datasets for medical LLMs
|
as the quality of these datasets directly impacts the models
|
medical capabilities Thus we analyze the training datasets
|
for existing medical LLMs from two perspectives the source
|
of the training corpus and data processing methods Training
|
corpus sources can assist developers in collecting the expected
|
data while data processing methods facilitate the cleansing
|
of data to form a standardized training set Furthermore
|
considering the differing capabilities injected to medical LLMs
|
by the three training stages of CP IFT and HA as well
|
as the variations in required computational resources and
|
data scale we have summarized the combinations of training
|
stages adopted by existing medical LLMs and classified them
|
into four paradigms IFT CPIFT IFTHA and CPIFT
|
HA These paradigms can guide institutions in choosing the
|
appropriate training methods based on their computational
|
capabilities data scale and specific needs Subsequently
|
given the professional nature of content generated by medical
|
LLMs and the potential harm that erroneous content can
|
cause to patients we systematically review existing evaluation
|
benchmarks and protocols from two perspectives machine
|
evaluation and humancentric evaluation In particular we
|
standardize the dimensions of the humancentric evaluationFinally we analyze the shortcomings of existing medical
|
LLMs and propose future research directions with the potential
|
to address these issues
|
Here we compare several recent similar studies The study
|
20 systematically summarizes the training data methods
|
optimization strategies and evaluation techniques for LLMs
|
in healthcare However our work retrieved updated literature
|
on medical LLMs and summarized the training data and
|
methods from a more nuanced perspective Specifically we
|
analyzed the data sources and processing techniques in detail
|
for training data For training methods we identified four
|
training paradigms and examined their respective advantages
|
and disadvantages Another study similar to ours is 21
|
which systematically overviews the methods datasets and
|
applications details of language foundation models vision
|
foundation models bioinformatics foundation models and
|
multimodal foundation models in healthcare Our work specif
|
ically focuses on language foundation models summarizing
|
datasets training methods and evaluation benchmarks using
|
new classifications and perspectives Finally some studies
|
19 22 have concentrated on the applications and challenges
|
of LLMs in medicine In contrast our research specifically
|
emphasizes methods and technologies
|
The contributions of this review are summarized below
|
1 We systematically review medical LLMs emerged since
|
ChatGPT Detailed guidelines and tutorials are provided
|
for healthcare organizations in training customized med
|
ical LLMs in terms of dataset construction trainingJOURNAL OF L ATEX CLASS FILES VOL 14 NO 8 AUGUST 2021 3
|
technologies and evaluation benchmarks
|
2 Data Acquisition and Processing Section II Summa
|
rizing and classifying existing training corpus sources
|
and data processing methods for medical LLMs offers
|
robust recommendations and guidance for constructing
|
customized medical datasets
|
3 Training Paradigms Section III Systematically sum
|
marizes the existing training methods for medical LLMs
|
and proposes four distinct training paradigms This fa
|
cilitates medical organizations in selecting appropriate
|
training methods based on their available computational
|
resources
|
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.