tensorkelechi/arxiv-cstext · Datasets at Hugging Face

text stringlengths 0 89.3k
1 Data Cleaning Data cleaning primarily aims to remove
noise and correct errors and inconsistencies thereby improving
overall data quality We have deeply investigated how the
existent medical LLM work cleans the data This process can
be categorized by function into Quality Filtering Sensitive
Information Filtering and Deduplication
Quality Filtering enhances data quality by using keywords
or special tags to clean the target data This process can be
implemented in various ways with keyword filtering being a
primary method Keyword filtering helps identify and retain
relevant medical domain data removing nonmedical content
97 By verifying whether a corpus fragment contains perti
nent medical keywords we can effectively determine if it be
longs to the desired data HuatuoGPTII 45 uses a dictionary
based method to determine if a text fragment belongs to the
target domain by evaluating the density of domainspecific
words Similarly Apollo 54 uses a dictionarybased approach
to select medical data from largescale Chinese English
and Spanish corpora MMedLM 49 employs a rulebased
filtering pipeline for extracting target data from multilingual
corpora It constructs a list of 1200 multilingual medical
terms and evaluates text passages based on the count andJOURNAL OF L ATEX CLASS FILES VOL 14 NO 8 AUGUST 2021 8
density of these terms considering a passage relevant only
if both metrics exceed a predefined threshold Additionally
nonessential sections of medical textbooks such as the cover
table of contents and end pages are excluded to maintain a
focus on medical content
Keyword filtering not only facilitates the extraction of
target domain data but also enhances data quality MedAlpaca
29 ensured data quality by selecting only forum replies
that received at least five votes Qibo 55 improved data
quality through characterlevel and paragraphlevel rules to
verify the plausibility of individual characters and the semantic
continuity of the text These rules were iteratively updated
and refined through sampling and manual checks InMDX 9
initially identified topranking journal titles within each inter
nal medicine subspecialty based on Journal Citation Reports
JCR ensuring data quality Subsequently it queried these
journals via the Pubmed API and extracted papers published
since 2010 to capture recent medical knowledge ChatDoctor
25 improved data quality by filtering out short meaningless
conversations BianQue 39 reduced data noise by removing
private content incomplete JSON data website links voice
recordings and autoresponders from medical consulting plat
form data using regular expressions Certain collected data
may include medical advertising with misleading language
potentially biasing medical LLMs HuatuoGPTII 45 utilizes
ChatGPT to identify advertisements within text and trains a
quality classification model to filter out such content from the
dataset
The collected data may contain errors potentially biasing
the models understanding and leading to inaccuracies in
generated text This could result in misinformation such as in
correct disease names which can confuse users Additionally
grammatical errors in the collected data may be reflected in
the models output reducing readability and user experience
To address this ChatDoctor manually filters erroneous content
and utilizes LanguageTool for grammatical corrections while
HuatuoGPT prompts LLMs to filter data for misspellings
slang and irrelevant information
The collected data may include special characters like
citation marks or URLs which lack contextual meaning
and are unsuitable for training medical LLMs PMCLLaMA
30 eliminates irrelevant elements from medical textbooks
such as URLs author lists redundant information document
content references citations and citation marks for images
and tables MediTron 44 removes URLs citations graphics
table separators and incorrectly formatted characters from
clinical guidelines In processing the literature author de
tails bibliographies acknowledgments tables and graphics
are excluded retaining only the main body of each paper
Additionally it employs to denote major sections and
for subsections replacing citation marks with the title of the
reference
Although manual review is timeconsuming and labor
intensive it remains the most reliable method for ensuring data
quality OncoGPT 53 submits automatically cleaned data for
review by domain experts and clinicians to correct inaccuracies
and ensure the authority of the answers
Sensitive Information Filtering Private information inmedical data usually refers to patients personal identity and
health information which can seriously harm patients rights
if disclosed Moreover including private attributes like race
gender and age may introduce bias in certain demographic
groups Common techniques for mitigating privacy concerns
include data desensitization substituting personal identifiers
for sensitive information within the text and data deletion
which involves directly removing sensitive details from the
text However the latter approach may compromise the in
tegrity and usability of the data ChiMedGPT and OncoGPT
remove private information including personal details hos
pital data and website information from dialog data This
process mitigates the impact of data redundancy and non
critical information on model performance
Deduplication removes duplicates reduces data redun
dancy and enhances the quality and diversity of training data
Additionally it prevents data leakage between the training and
test sets ensuring a fair evaluation process HuatuoGPTII
45 employs a sentence embedding model to transform the
corpus into embedding vectors and subsequently eliminates
semantically similar texts through an intensive search method

1 Data Cleaning Data cleaning primarily aims to remove

noise and correct errors and inconsistencies thereby improving

overall data quality We have deeply investigated how the

existent medical LLM work cleans the data This process can

be categorized by function into Quality Filtering Sensitive

Information Filtering and Deduplication

Quality Filtering enhances data quality by using keywords

or special tags to clean the target data This process can be

implemented in various ways with keyword filtering being a

primary method Keyword filtering helps identify and retain

relevant medical domain data removing nonmedical content

97 By verifying whether a corpus fragment contains perti

nent medical keywords we can effectively determine if it be

longs to the desired data HuatuoGPTII 45 uses a dictionary

based method to determine if a text fragment belongs to the

target domain by evaluating the density of domainspecific

words Similarly Apollo 54 uses a dictionarybased approach

to select medical data from largescale Chinese English

and Spanish corpora MMedLM 49 employs a rulebased

filtering pipeline for extracting target data from multilingual

corpora It constructs a list of 1200 multilingual medical

terms and evaluates text passages based on the count andJOURNAL OF L ATEX CLASS FILES VOL 14 NO 8 AUGUST 2021 8

density of these terms considering a passage relevant only

if both metrics exceed a predefined threshold Additionally

nonessential sections of medical textbooks such as the cover

table of contents and end pages are excluded to maintain a

focus on medical content

Keyword filtering not only facilitates the extraction of

target domain data but also enhances data quality MedAlpaca

29 ensured data quality by selecting only forum replies

that received at least five votes Qibo 55 improved data

quality through characterlevel and paragraphlevel rules to

verify the plausibility of individual characters and the semantic

continuity of the text These rules were iteratively updated

and refined through sampling and manual checks InMDX 9

initially identified topranking journal titles within each inter

nal medicine subspecialty based on Journal Citation Reports

JCR ensuring data quality Subsequently it queried these

journals via the Pubmed API and extracted papers published

since 2010 to capture recent medical knowledge ChatDoctor

25 improved data quality by filtering out short meaningless

conversations BianQue 39 reduced data noise by removing

private content incomplete JSON data website links voice

recordings and autoresponders from medical consulting plat

form data using regular expressions Certain collected data

may include medical advertising with misleading language

potentially biasing medical LLMs HuatuoGPTII 45 utilizes

ChatGPT to identify advertisements within text and trains a

quality classification model to filter out such content from the

dataset

The collected data may contain errors potentially biasing

the models understanding and leading to inaccuracies in

generated text This could result in misinformation such as in

correct disease names which can confuse users Additionally

grammatical errors in the collected data may be reflected in

the models output reducing readability and user experience

To address this ChatDoctor manually filters erroneous content

and utilizes LanguageTool for grammatical corrections while

HuatuoGPT prompts LLMs to filter data for misspellings

slang and irrelevant information

The collected data may include special characters like

citation marks or URLs which lack contextual meaning

and are unsuitable for training medical LLMs PMCLLaMA

30 eliminates irrelevant elements from medical textbooks

such as URLs author lists redundant information document

content references citations and citation marks for images

and tables MediTron 44 removes URLs citations graphics

table separators and incorrectly formatted characters from

clinical guidelines In processing the literature author de

tails bibliographies acknowledgments tables and graphics

are excluded retaining only the main body of each paper

Additionally it employs to denote major sections and

for subsections replacing citation marks with the title of the

reference

Although manual review is timeconsuming and labor

intensive it remains the most reliable method for ensuring data

quality OncoGPT 53 submits automatically cleaned data for

review by domain experts and clinicians to correct inaccuracies

and ensure the authority of the answers

Sensitive Information Filtering Private information inmedical data usually refers to patients personal identity and

health information which can seriously harm patients rights

if disclosed Moreover including private attributes like race

gender and age may introduce bias in certain demographic

groups Common techniques for mitigating privacy concerns

include data desensitization substituting personal identifiers

for sensitive information within the text and data deletion

which involves directly removing sensitive details from the

text However the latter approach may compromise the in

tegrity and usability of the data ChiMedGPT and OncoGPT

remove private information including personal details hos

pital data and website information from dialog data This

process mitigates the impact of data redundancy and non

critical information on model performance

Deduplication removes duplicates reduces data redun

dancy and enhances the quality and diversity of training data

Additionally it prevents data leakage between the training and

test sets ensuring a fair evaluation process HuatuoGPTII

45 employs a sentence embedding model to transform the

corpus into embedding vectors and subsequently eliminates

semantically similar texts through an intensive search method