LLM fine-tuning

with medical records 이론/실습

2025-10-23, 15:00 ~ 17:00

Seongsu Bae, Sujeong Im

KAIST AI @ Edlab (Advised by Edward Choi)

KoSAIM 2025 개발자를 위한 AI 실습교육: Train your own medical AI

Speaker Bio

Sujeong Im (임수정)

Education

• POSTECH Creative IT Engineering, B.Sc�(2018-2022)
• KAIST Kim Jaechul Graduate School of AI, M.Sc (2023-2025)
• KAIST Kim Jaechul Graduate School of AI, Ph.D (2025-)

Research Interests

• Foundation Model
• Natural Language Processing
• Machine Learning for Healthcare

Seongsu Bae (배성수)

Education

• Hanyang University Mathematics, B.Sc (2013-2019)
• KAIST Kim Jaechul Graduate School of AI, M.Sc (2020-2022)
• KAIST Kim Jaechul Graduate School of AI, Ph.D (2022-)

Research Interests

• Semantic Machine
• Multimodal Learning
• Machine Learning for Healthcare

Table of Contents

• How to build a clinical domain Large Language Model (LLM)? (40 mins)
• (Large) Language Model
• How to build a (large) language model?
• Building an instruction-following LLM in the clinical domain
• Asclepius (Gweon and Kim et al., ACL 2024 Findings)�
• Hands-on Session: Fine-tuning a clinical domain LLM (80 mins)
• Environment Setup & Colab Practice
• LLM memory layout
• Parameter-Efficient Fine-Tuning (LoRA/QLoRA)

Language Model

We deal with LMs every day!

How to train a LM?

The

The

sky

is

blue

.

The

sky

is

The

sky

blue

is

sky

Next Token Prediction task for the sentence “The sky is blue.”

Text Generation via a Probabilistic Model

The

sky

is

blue

clear

usually

the

<

(Large)

Language Model

more likely

less likely

How to build a (large) language model?

• Pre-training and Fine-tuning
• e.g., BERT (2018), T5 (2019)

Pretrained

LM

Finetune on task A

Finetune on task B

Finetune on task C

Inference

on task A

Inference

on task B

Inference

on task C

(-) Task-specific training → One specialized model for each task

How to build a (large) language model?

• Pre-training and Prompting
• e.g., GPT-3 (2020)

Pretrained

LM

Inference

on task A

Inference

on task B

Inference

on task C

(+) Improve performance via few-shot prompting or prompt engineering

prompting

How to build a (large) language model?

• Pre-training and Prompting

(-) Forced few-shot prompting

(-) Manual efforts for the prompting technique

(-) Not aligned with natural instructions

How to build a (large) language model?

• Pre-training and Instruction tuning
• Supervised Fine-Tuning (SFT) on instruction data
• e.g., FLAN (2021), LLaMA (2023)

Pretrained

LM

Inference

on task A

Inference

on task B

Inference

on task C

(+) model learns to perform many tasks via natural language instructions

…

instructions

fine-tune on many instructions

How to build a (large) language model?

• Pre-training and Alignment tuning
• Supervised Fine-Tuning (SFT) on instruction data�+ Alignment learning on preference data (e.g., RLHF, DPO)
• e.g., InstructGPT (2022), ChatGPT (2022), Llama 2 (2023), Llama 3 (2024)

Building an instruction-following LLM

• How can we build an instruction-following LLM?
• Prepare a pre-trained large language model (e.g., LLaMA 7B)
• Perform supervised fine-tuning on instruction data (e.g., Alpaca 52K dataset)��
• How can we build an instruction-following LLM in the clinical domain?
• Prepare a pre-trained large language model
• Pre-training on clinical corpus for domain adaptation
• Perform supervised fine-tuning using domain-specific clinical instruction data
• Today, we will focus on instruction-following data tailored for clinical notes!

Imagine a clinical LLM

• Given a clinical note, a clinical LLM can perform these tasks as follows:
• “What medical procedures were performed on the patient during her hospital course, as mentioned in the discharge summary?” Named Entity Recognition
• “What abbreviation was expanded using the acronym ‘ANH’ in the diagnosis section of the discharge summary?” Abbreviation Expansion
• “When was the patient started on oral acyclovir and what was the duration of treatment?” Temporal Information Extraction
• “Can you summarize the patient’s hospital course, treatment, and diagnoses according to the given discharge summary?” Summarization
• “What was the reason for the patient’s transfer to ICU and what was the treatment plan for infection-induced respiratory failure?” Question Answering

Asclepius: Publicly Shareable Clinical Large Language Model Built on Synthetic Clinical Notes (Gweon and Kim et al., ACL 2024 Findings)

Real clinical note

• Semi-Structured Text about Patient Activity
• Properties
• Semi-structured: Associated with headers
• Acronyms
• Typos
• Problem: Protected Health Information (PHI)
• Use GPT: PHI ⇒ Impractical
• Human Annotation: Require Experts ⇒ cost
• Machine Annotation: PHI ⇒ Impractical

Case report

• To share “case” with community
• No PHI ⇒ Sharable
• Properties
• Plain text
• Less acronyms
• Well-written
• Contents are similar to the notes
• e.g., PMC (PubMed Central) case report

Synthetic clinical note generation

Clinical instruction/response data generation

Final dataset

• (clinical note, instruction, response) triples ⇒ all synthetics!

Asclepius-Llama3-8B

• How can we build an instruction-following LLM in the clinical domain?
• Prepare a pre-trained large language model
• use Llama3-8B model
• Pre-training on clinical corpus for domain adaptation
• Pre-training (1 epoch): 2h 59m with 4x A100 80G
• dataset: synthetic clinical notes
• Perform supervised fine-tuning using domain-specific clinical instruction data
• Instruction fine-tuning (3 epoch): 30h 41m with 4x A100 80G
• dataset: clinical instruction-response pairs with synthetic clinical notes

Hands-on Session:

Fine-tuning a clinical domain LLM

Environment Setup

• https://github.com/baeseongsu/

colab link

Environment Setup

• https://github.com/baeseongsu/Clinical-LLM-FineTuning-HandsOn

Environment Setup

Colab Objectives

• Goal: Fine-tuning a clinical domain LLM
• Environment: Google Colab
• Dataset: starmpcc/Asclepius-Synthetic-Clinical-Notes
• Model: microsoft/phi-2 (2.7B)
• CAUTION (주의)
• LLM 학습하는 과정에서 Colab을 절대 끄지 마시기 바랍니다.
• 새로고침 금지
• 코랩 내에서 다른 버튼 클릭 금지
• 실행 중지 금지

Deep learning memory layout

• Model size: B (billion) scale
• xB parameters = xB floating point numbers = 2x GB (bf16/fp16)
• Deep Learning Memory Requirements
• model parameter: 2x GB
• gradient state: 2x GB
• optimizer state: 2x ~ 12x GB
• Total: 6~16x GB + alpha
• Our requirements
• model: phi-2 (2.7B)
• GPU VRAM: Colab T4 (16GB)
• 2.7*6=16.2

Can You Run it?

• https://huggingface.co/spaces/Vokturz/can-it-run-llm

LoRA (Hu and Shen et al., 2021)

QLoRA (Dettmers and Pagnoni et al., 2023)

Parameter-Efficient Fine-Tuning (PEFT)

• https://github.com/huggingface/peft

Thank you :D

If you require any further information, feel free to contact us: seongsu@kaist.ac.kr, sujeongim@kaist.ac.kr

KoSAIM 2025 개발자를 위한 AI 실습교육: Train your own medical AI