tensorkelechi/arxiv-cstext · Datasets at Hugging Face

text stringlengths 0 89.3k
of medical problems facilitating the generation of reasoning
steps and answers AntGLMMed 46 introduces a simple
prompt strategy called VerificationofChoice V oC tailored
for the multiplechoice QA benchmark V oC operates underthe assumption that each choice is correct prompting the LLM
to provide the corresponding explanation Subsequently these
explanations serve as context for the LLM to identify the
inconsistencies and give the final answer
Fewshot learning evaluates the performance of LLMs on
downstream tasks involving a limited number of examples
Typically examples are derived from the training set in the
downstream task including both inputs and corresponding
outputs The medical LLM then utilizes these examples in
conjunction with its intrinsic medical knowledge to gen
erate appropriate outputs in accordance with the provided
instructions Beyond simply employing a limited number of
examples as context several studies 44 have incorporated
prompting strategies like ChainofThought 135 and Self
Consistency ChainofThought SCCoT 136 to stimulate
the reasoning capabilities of medical LLMs These strategies
aim to enhance the accuracy of outputs and the rationality
of reasoning processes Additionally prompting strategies can
facilitate the identification and correction of potential errors
and knowledge gaps in medical LLMs particularly in complex
medical tasks
Taskspecific supervised finetuning involves training a
medical LLM on a benchmark training set and evaluating it
on a corresponding test set 41 48 By training the medical
LLM on these specific medical tasks such as drug named
entity recognition and medical literature comprehension the
generalization performance and knowledge transfer ability of
the model can be evaluated Two types of training approaches
parameterefficient finetuning and fullparameter finetuning
are used to evaluate the performance of the model under two
different levels of medical knowledge injections 49
B HumanCentric Evaluation
Given the inherent diversity and complexity of model
responses machine evaluation methods may not adequately
capture the performance of medical LLMs particularly in
assessing the usefulness and safety of outputs To address
this limitation researchers have increasingly adopted Human
Centric Evaluation approaches It provides a more realistic
assessment of medical LLMs in realworld applications re
flecting human judgment and expectations
1 Evaluator Existence studies employ both humans and
LLMs as evaluators Since the users of medical LLMs are
humans human evaluation can be effective and direct in
assessing medical LLMs Given the specialized nature of the
medical field existing work categorizes human evaluation
into expert evaluation and lay user evaluation reflecting the
perspectives of physicians and general users respectively
While human evaluation offers valuable insights into LLM per
formance it is not without its limitations Human evaluation
can be timeconsuming and resourceintensive and the results
can be susceptible to individual biases and expertise levels
To address these challenges some researchers have explored
employing advanced LLMs to replace human evaluation of
medical LLMs These advanced LLMs are trained on vast
datasets and possess a knowledge base comparable to that
of humans Since the LLM is to evaluate the performance ofJOURNAL OF L ATEX CLASS FILES VOL 14 NO 8 AUGUST 2021 16
medical LLMs from a human perspective the LLM evaluation
is included in HumanCentric Evaluation
2 Evaluation Dimensions In HumanCentric Evaluation
the initial step involves identifying the evaluation dimensions
for the model Given the inconsistent terminology employed
in previous studies of medical LLMs to describe these dimen
sions we standardized the terminology and provided corre
sponding explanations Tab III presents the normalized terms
Doctorlike Exhibits the communication style and tone of
a professional physician reflecting a comprehensive under
standing of medical principles and clinical expertise Capable
of providing clear explanations of intricate medical concepts
and offering wellfounded diagnostic insights and recommen
dations
Patientfriendly Interacts with users in a patient and
empathetic manner fostering a sense of understanding and
support Effectively simplifies complex medical concepts into
easily comprehensible language demonstrating compassion
and care thereby alleviating user anxiety and enhancing trust
and satisfaction
Professional Exhibits a high degree of professionalism
characterized by comprehensive medical knowledge and sound
clinical judgment Provides treatment recommendations and
prescriptions that are not only accurate and reliable but
are also conform to medical norms and ethics ensuring the
achievement of optimal patient outcomes
Safe Responses are safe and harmless and not geograph
ically or racially discriminatory Correctly address unhealthy
and unsafe issues provide accurate and reliable advice and
avoid misleading users or raising health risks Always follow
medical ethics protect user privacy and ensure the accuracy
and safety of information delivery
Fluent Responses should be fluent coherent and logical
to ensure readability and comprehension Accurate medical
terminology should be employed without overspecialization
conveying the message professionally and understandably
Proactive Capable of proactively asking users to add
relevant information and ask targeted additional questions
when faced with unclear descriptions of symptoms or lack
of necessary information
Helpful Capable of meeting the users needs for medical

of medical problems facilitating the generation of reasoning

steps and answers AntGLMMed 46 introduces a simple

prompt strategy called VerificationofChoice V oC tailored

for the multiplechoice QA benchmark V oC operates underthe assumption that each choice is correct prompting the LLM

to provide the corresponding explanation Subsequently these

explanations serve as context for the LLM to identify the

inconsistencies and give the final answer

Fewshot learning evaluates the performance of LLMs on

downstream tasks involving a limited number of examples

Typically examples are derived from the training set in the

downstream task including both inputs and corresponding

outputs The medical LLM then utilizes these examples in

conjunction with its intrinsic medical knowledge to gen

erate appropriate outputs in accordance with the provided

instructions Beyond simply employing a limited number of

examples as context several studies 44 have incorporated

prompting strategies like ChainofThought 135 and Self

Consistency ChainofThought SCCoT 136 to stimulate

the reasoning capabilities of medical LLMs These strategies

aim to enhance the accuracy of outputs and the rationality

of reasoning processes Additionally prompting strategies can

facilitate the identification and correction of potential errors

and knowledge gaps in medical LLMs particularly in complex

medical tasks

Taskspecific supervised finetuning involves training a

medical LLM on a benchmark training set and evaluating it

on a corresponding test set 41 48 By training the medical

LLM on these specific medical tasks such as drug named

entity recognition and medical literature comprehension the

generalization performance and knowledge transfer ability of

the model can be evaluated Two types of training approaches

parameterefficient finetuning and fullparameter finetuning

are used to evaluate the performance of the model under two

different levels of medical knowledge injections 49

B HumanCentric Evaluation

Given the inherent diversity and complexity of model

responses machine evaluation methods may not adequately

capture the performance of medical LLMs particularly in

assessing the usefulness and safety of outputs To address

this limitation researchers have increasingly adopted Human

Centric Evaluation approaches It provides a more realistic

assessment of medical LLMs in realworld applications re

flecting human judgment and expectations

1 Evaluator Existence studies employ both humans and

LLMs as evaluators Since the users of medical LLMs are

humans human evaluation can be effective and direct in

assessing medical LLMs Given the specialized nature of the

medical field existing work categorizes human evaluation

into expert evaluation and lay user evaluation reflecting the

perspectives of physicians and general users respectively

While human evaluation offers valuable insights into LLM per

formance it is not without its limitations Human evaluation

can be timeconsuming and resourceintensive and the results

can be susceptible to individual biases and expertise levels

To address these challenges some researchers have explored

employing advanced LLMs to replace human evaluation of

medical LLMs These advanced LLMs are trained on vast

datasets and possess a knowledge base comparable to that

of humans Since the LLM is to evaluate the performance ofJOURNAL OF L ATEX CLASS FILES VOL 14 NO 8 AUGUST 2021 16

medical LLMs from a human perspective the LLM evaluation

is included in HumanCentric Evaluation

2 Evaluation Dimensions In HumanCentric Evaluation

the initial step involves identifying the evaluation dimensions

for the model Given the inconsistent terminology employed

in previous studies of medical LLMs to describe these dimen

sions we standardized the terminology and provided corre

sponding explanations Tab III presents the normalized terms

Doctorlike Exhibits the communication style and tone of

a professional physician reflecting a comprehensive under

standing of medical principles and clinical expertise Capable

of providing clear explanations of intricate medical concepts

and offering wellfounded diagnostic insights and recommen

dations

Patientfriendly Interacts with users in a patient and

empathetic manner fostering a sense of understanding and

support Effectively simplifies complex medical concepts into

easily comprehensible language demonstrating compassion

and care thereby alleviating user anxiety and enhancing trust

and satisfaction

Professional Exhibits a high degree of professionalism

characterized by comprehensive medical knowledge and sound

clinical judgment Provides treatment recommendations and

prescriptions that are not only accurate and reliable but

are also conform to medical norms and ethics ensuring the

achievement of optimal patient outcomes

Safe Responses are safe and harmless and not geograph

ically or racially discriminatory Correctly address unhealthy

and unsafe issues provide accurate and reliable advice and

avoid misleading users or raising health risks Always follow

medical ethics protect user privacy and ensure the accuracy

and safety of information delivery

Fluent Responses should be fluent coherent and logical

to ensure readability and comprehension Accurate medical

terminology should be employed without overspecialization

conveying the message professionally and understandably

Proactive Capable of proactively asking users to add

relevant information and ask targeted additional questions

when faced with unclear descriptions of symptoms or lack

of necessary information

Helpful Capable of meeting the users needs for medical