tensorkelechi/arxiv-cstext · Datasets at Hugging Face

text stringlengths 0 89.3k
4 Medical LLMs Evaluations Section IV Summarizes
the existing benchmarks for evaluating medical LLMs
and categorizes them into machine and humancentric
evaluations Additionally the humancentric evaluation
dimensions of medical LLMs are normalized This
provides a comprehensive assessment perspective and
methodology for healthcare organizations
5 Challenges and Future Directions Section V Future
research directions are proposed to address the short
comings and research gaps of the existing medical
LLMs This provides important guidance and insights
to further exploration in Medical LLMs for medical
organizations and researchers
II D ATA ACQUISITION AND PROCESSING
Datasets are crucial for transferring knowledge of LLMs
from general domains to specific ones The quality of a
dataset significantly impacts model performance A high
quality dataset provides accurate supervised signals enabling
the model to effectively learn domainspecific rules and pat
terns This not only improves the models alignment with the
target task but also break the scaling law to some extent
23 Medical LLMs should be doctorlike patientfriendly
and professional 8 To ensure that these models are able
to learn the language characteristics and terminology of the
medical domain training datasets should meet the above
guidelines To construct a medical domain dataset that meets
the requirements of a specific task it is essential to identify
appropriate training corpus sources These sources may vary
in their medical discipline focus level of specialization and
data quality Therefore it is crucial to select a corpus source
that meets the needs After identifying a suitable corpus
source the corpus cannot be directly used for model training
They may contain inappropriate content such as violent
pornographic or discriminatory material or have data format
irregularities Therefore data processing is required to ensure
the corpus meets the necessary training requirements The
training of medical LLMs currently requires three types of
data unstructured data for continued pretraining question
answer pairs both single and multiple turns for instruction
finetuning and human preference data for human alignment
Here we systematically summarize the training corpus sources
and data processing methods for investigated medical LLMs in
Tab I providing a reference for constructing custom medical
training sets In addition we categorized the corpus sources
and data processing methods as shown in Fig 2
Existing Public DatasetsTraining Corpus Sources
Unstructured Human Preference
NLP TaskSynthetic Data
QA
Public Medical Corpus
Encyclopedia Textbook
Package InsertGuideline
Medical Website
Medical Academic LiteratureClinical NoteProfessional Medical
Organization Corpus
Electronic Health Record
Medical Prescription
Doctor Patient
Conversation
Data Processes
Data Cleaning Data Formatting
ConstructionQuality Filtering
Sensitive Information
Filtering
DeduplicationUnstructured Data
Instruction Data
Human Preference Data
Data Augmentation
Unified Expression
EnrichmentTranslation
Accuracy
Localization
Unstructured Data
Instruction Data
Human Preference DataTraining SetsKG QAFig 2 Detailed Categorization of Corpus Sources and Data Processing
Methods
A Training Corpus Sources
We categorized the corpus sources into four groups Existing
Public Datasets Public Medical Corpus Professional Medical
Organization Corpus and Synthetic Data The strengths and
weaknesses of each category as well as their suitability
for different training phases of medical LLMs are detailed
below Additionally guidance is provided for customizing the
collection of training data
1 Existing Public Datasets It is recommended to use
existing public datasets for constructing customized training
sets for medical LLMs These datasets are welldesigned high
quality and widely recognized in existing research The public
datasets are further categorized into five types unstructured
data questionanswer pairs human preference data knowl
edge graphs and regular NLP task data Related work 20
21 have systematically summarized datasets for LLMs in the
medical domain Here we focus on the characteristics of each
type of data and considerations for using these public datasets
Unstructured Data usually refers to plain text data such as
the numerous public corpora including CommonCrawl 56
C4 57 Wikipedia 58 Baidu Baike 59 UFAL Medical
Corpus 60 and other LLM collections like RedPajama 61
These corpora are collected from web data using crawlers re
sulting in a largescale dataset However they require rigorous

4 Medical LLMs Evaluations Section IV Summarizes

the existing benchmarks for evaluating medical LLMs

and categorizes them into machine and humancentric

evaluations Additionally the humancentric evaluation

dimensions of medical LLMs are normalized This

provides a comprehensive assessment perspective and

methodology for healthcare organizations

5 Challenges and Future Directions Section V Future

research directions are proposed to address the short

comings and research gaps of the existing medical

LLMs This provides important guidance and insights

to further exploration in Medical LLMs for medical

organizations and researchers

II D ATA ACQUISITION AND PROCESSING

Datasets are crucial for transferring knowledge of LLMs

from general domains to specific ones The quality of a

dataset significantly impacts model performance A high

quality dataset provides accurate supervised signals enabling

the model to effectively learn domainspecific rules and pat

terns This not only improves the models alignment with the

target task but also break the scaling law to some extent

23 Medical LLMs should be doctorlike patientfriendly

and professional 8 To ensure that these models are able

to learn the language characteristics and terminology of the

medical domain training datasets should meet the above

guidelines To construct a medical domain dataset that meets

the requirements of a specific task it is essential to identify

appropriate training corpus sources These sources may vary

in their medical discipline focus level of specialization and

data quality Therefore it is crucial to select a corpus source

that meets the needs After identifying a suitable corpus

source the corpus cannot be directly used for model training

They may contain inappropriate content such as violent

pornographic or discriminatory material or have data format

irregularities Therefore data processing is required to ensure

the corpus meets the necessary training requirements The

training of medical LLMs currently requires three types of

data unstructured data for continued pretraining question

answer pairs both single and multiple turns for instruction

finetuning and human preference data for human alignment

Here we systematically summarize the training corpus sources

and data processing methods for investigated medical LLMs in

Tab I providing a reference for constructing custom medical

training sets In addition we categorized the corpus sources

and data processing methods as shown in Fig 2

Existing Public DatasetsTraining Corpus Sources

Unstructured Human Preference

NLP TaskSynthetic Data

QA

Public Medical Corpus

Encyclopedia Textbook

Package InsertGuideline

Medical Website

Medical Academic LiteratureClinical NoteProfessional Medical

Organization Corpus

Electronic Health Record

Medical Prescription

Doctor Patient

Conversation

Data Processes

Data Cleaning Data Formatting

ConstructionQuality Filtering

Sensitive Information

Filtering

DeduplicationUnstructured Data

Instruction Data

Human Preference Data

Data Augmentation

Unified Expression

EnrichmentTranslation

Accuracy

Localization

Unstructured Data

Instruction Data

Human Preference DataTraining SetsKG QAFig 2 Detailed Categorization of Corpus Sources and Data Processing

Methods

A Training Corpus Sources

We categorized the corpus sources into four groups Existing

Public Datasets Public Medical Corpus Professional Medical

Organization Corpus and Synthetic Data The strengths and

weaknesses of each category as well as their suitability

for different training phases of medical LLMs are detailed

below Additionally guidance is provided for customizing the

collection of training data

1 Existing Public Datasets It is recommended to use

existing public datasets for constructing customized training

sets for medical LLMs These datasets are welldesigned high

quality and widely recognized in existing research The public

datasets are further categorized into five types unstructured

data questionanswer pairs human preference data knowl

edge graphs and regular NLP task data Related work 20

21 have systematically summarized datasets for LLMs in the

medical domain Here we focus on the characteristics of each

type of data and considerations for using these public datasets

Unstructured Data usually refers to plain text data such as

the numerous public corpora including CommonCrawl 56

C4 57 Wikipedia 58 Baidu Baike 59 UFAL Medical

Corpus 60 and other LLM collections like RedPajama 61

These corpora are collected from web data using crawlers re

sulting in a largescale dataset However they require rigorous