tensorkelechi/arxiv-cstext · Datasets at Hugging Face

text stringlengths 0 89.3k
paradigm is that it requires more training data extensive com
putational resources and longer training times compared to
the IFT paradigm Additionally if the data used for continued
pretraining is irrelevant or significantly different from the
instruction finetuning tasks the model may learn irrelevant
information which can degrade performance on downstream
tasks
1 Model Merging Model fusion refers to the merging of
two or more models in order to improve the generalization
performance of the model without additional training Most
methods focus on fusing model parameters BioMistral merges
the pretrained medical LLMs with the base model using three
methods TIES DARE and SLERP for weight merging
respectively to improve the general domain capabilities of
the medical LLMs Apollo 54 uses Proxy Tuning 124 to
indirectly guide the adjustment of general LLMs with a largenumber of parameters by utilizing the logits outputs of the
prefinetuning model without medical knowledge injected
and the finetuned model with medical knowledge injected
Experimental findings indicate that proxy tuning enhances
the models performance on multiple language benchmarks
beyond the English benchmark
2 Combining Pretraining Data with Instruction Data
Incorporating pretraining data into instruction finetuning by
converting pretraining data into the format of instruction data
and jointly training it with other instruction data enhances
the effectiveness and stability of training 20 Building upon
this concept HuatuoGPTII 45 transforms its pretrained
medical data into an instruction format creating a pretrained
instruction dataset This dataset is then jointly trained with
existing instruction data Additionally it proposes a priority
sampling strategy to mitigate the challenges associated with
mixing data sources during LLM training Similarly Apollo
54 rewrites the pretraining corpus into the data format of QA
and adopts the same data mixing and sampling techniques as
HuatuoGPTII This approach facilitates a smoother transition
from continued pretraining to instruction finetuning
C IFTHA
The IFTHA paradigm adopts a twostage approach In the
first stage instruction finetuning is employed to augment the
LLMs ability to follow medical instructions and engage in
medical dialogue Subsequently the second stage focuses on
human alignment ensuring the LLMs responses adhere to
established medical alignment principles In general domain
human alignment aims to ensure that LLMs operate in accor
dance with human values genuine intentions and established
social ethics The main alignment principles are helpful
honest and harmless Within the medical domain these core
principles are further extended to encompass patientFridendly
and DoctorLike professional and interactive diagnostic 8
To achieve human alignment in the medical domain a
prevalent approach involves Reinforcement Learning from
Human Feedback RLHF 14 This methodology entails a
twostage process Initially a reward model is constructed
on top of an instruction finetuned medical LLM utilizing
manually labeled human preference data Subsequently this
reward model is combined with either PPO 125 or Rejection
Sampling strategies 101 to perform reinforcement learning
training thereby aligning medical LLMs with predetermined
principles Considering the difficulty of collecting human
preference data HuatuoGPT motivated by RLHF and RLAIF
126 proposes Reinforcement Learning with Mixed Feedback
RLMF RLMF leverages ChatGPT to generate a portion ofJOURNAL OF L ATEX CLASS FILES VOL 14 NO 8 AUGUST 2021 13
the human preference data effectively merging the strengths
of ChatGPT and real doctors Consequently medical LLMs
are guided to align responses with both ChatGPT and medical
expertise QilinMed 41 adopts Direct Preference Optimiza
tion DPO 127 a nonreinforcement learning approach
to accomplish human alignment training for medical LLMs
Unlike reinforcement learning methods that rely on reward
modeling DPO directly establishes the connection between
the decision function and the reward function obviating the
need for explicit reward modeling
Currently human alignment training is built on models that
have undergone instruction finetuning The IFTHA paradigm
offers several advantages medical LLMs exhibit enhanced
instruction following capabilities yielding more accurate re
sponses to user queries and physician instructions Addition
ally IFTHA fosters alignment with human values promoting
the provision of accurate and reliable medical information
clear explanations and the avoidance of harmful sugges
tions while safeguarding user privacy However the IFTHA
paradigm presents challenges The human alignment train
ing process is intricate susceptible to instability and highly
sensitive to hyperparameter selection Furthermore RLHF is
highly dependent on the performance of the reward model
which in turn is dependent on the quality of the expensive
human preference data Consequently developers opting for
this paradigm need substantial engineering experience
D CPIFTHA
The CPIFTHA paradigm injects medical knowledge at
different dimensions into the general LLM from three stages
continued pretraining instruction finetuning and human align
ment Due to the complexity of training and large compu
tational resources required for this paradigm only a few
medical LLM efforts have adopted this paradigm However
the medical knowledge comprehension medical instruction
following and human preference abilities of medical LLMs
are improved compared to other paradigms
Compared to the CPIFT paradigm the paradigm intro
duces the human alignment phase This enhancement empow

paradigm is that it requires more training data extensive com

putational resources and longer training times compared to

the IFT paradigm Additionally if the data used for continued

pretraining is irrelevant or significantly different from the

instruction finetuning tasks the model may learn irrelevant

information which can degrade performance on downstream

tasks

1 Model Merging Model fusion refers to the merging of

two or more models in order to improve the generalization

performance of the model without additional training Most

methods focus on fusing model parameters BioMistral merges

the pretrained medical LLMs with the base model using three

methods TIES DARE and SLERP for weight merging

respectively to improve the general domain capabilities of

the medical LLMs Apollo 54 uses Proxy Tuning 124 to

indirectly guide the adjustment of general LLMs with a largenumber of parameters by utilizing the logits outputs of the

prefinetuning model without medical knowledge injected

and the finetuned model with medical knowledge injected

Experimental findings indicate that proxy tuning enhances

the models performance on multiple language benchmarks

beyond the English benchmark

2 Combining Pretraining Data with Instruction Data

Incorporating pretraining data into instruction finetuning by

converting pretraining data into the format of instruction data

and jointly training it with other instruction data enhances

the effectiveness and stability of training 20 Building upon

this concept HuatuoGPTII 45 transforms its pretrained

medical data into an instruction format creating a pretrained

instruction dataset This dataset is then jointly trained with

existing instruction data Additionally it proposes a priority

sampling strategy to mitigate the challenges associated with

mixing data sources during LLM training Similarly Apollo

54 rewrites the pretraining corpus into the data format of QA

and adopts the same data mixing and sampling techniques as

HuatuoGPTII This approach facilitates a smoother transition

from continued pretraining to instruction finetuning

C IFTHA

The IFTHA paradigm adopts a twostage approach In the

first stage instruction finetuning is employed to augment the

LLMs ability to follow medical instructions and engage in

medical dialogue Subsequently the second stage focuses on

human alignment ensuring the LLMs responses adhere to

established medical alignment principles In general domain

human alignment aims to ensure that LLMs operate in accor

dance with human values genuine intentions and established

social ethics The main alignment principles are helpful

honest and harmless Within the medical domain these core

principles are further extended to encompass patientFridendly

and DoctorLike professional and interactive diagnostic 8

To achieve human alignment in the medical domain a

prevalent approach involves Reinforcement Learning from

Human Feedback RLHF 14 This methodology entails a

twostage process Initially a reward model is constructed

on top of an instruction finetuned medical LLM utilizing

manually labeled human preference data Subsequently this

reward model is combined with either PPO 125 or Rejection

Sampling strategies 101 to perform reinforcement learning

training thereby aligning medical LLMs with predetermined

principles Considering the difficulty of collecting human

preference data HuatuoGPT motivated by RLHF and RLAIF

126 proposes Reinforcement Learning with Mixed Feedback

RLMF RLMF leverages ChatGPT to generate a portion ofJOURNAL OF L ATEX CLASS FILES VOL 14 NO 8 AUGUST 2021 13

the human preference data effectively merging the strengths

of ChatGPT and real doctors Consequently medical LLMs

are guided to align responses with both ChatGPT and medical

expertise QilinMed 41 adopts Direct Preference Optimiza

tion DPO 127 a nonreinforcement learning approach

to accomplish human alignment training for medical LLMs

Unlike reinforcement learning methods that rely on reward

modeling DPO directly establishes the connection between

the decision function and the reward function obviating the

need for explicit reward modeling

Currently human alignment training is built on models that

have undergone instruction finetuning The IFTHA paradigm

offers several advantages medical LLMs exhibit enhanced

instruction following capabilities yielding more accurate re

sponses to user queries and physician instructions Addition

ally IFTHA fosters alignment with human values promoting

the provision of accurate and reliable medical information

clear explanations and the avoidance of harmful sugges

tions while safeguarding user privacy However the IFTHA

paradigm presents challenges The human alignment train

ing process is intricate susceptible to instability and highly

sensitive to hyperparameter selection Furthermore RLHF is

highly dependent on the performance of the reward model

which in turn is dependent on the quality of the expensive

human preference data Consequently developers opting for

this paradigm need substantial engineering experience

D CPIFTHA

The CPIFTHA paradigm injects medical knowledge at

different dimensions into the general LLM from three stages

continued pretraining instruction finetuning and human align

ment Due to the complexity of training and large compu

tational resources required for this paradigm only a few

medical LLM efforts have adopted this paradigm However

the medical knowledge comprehension medical instruction

following and human preference abilities of medical LLMs

are improved compared to other paradigms

Compared to the CPIFT paradigm the paradigm intro

duces the human alignment phase This enhancement empow