Instruction Tuning

Creator
Creator
Seonglae ChoSeonglae Cho
Created
Created
2023 Jul 1 14:52
Editor
Edited
Edited
2024 Oct 16 0:12

Teach language models to follow instructions to solve a task

Instruction 데이터셋을 통해
Fine Tuning
을 진행하고 이를 통해
Zero shot learning
성능을 높이는 방법
Few Shot Prompting
과 달리 데이터를 직접적으로 보여주지 않아도 된다
산업적으로 데이터셋 누구나 쉽게 만들 수 있다는 게 핵심적 의의. Instruction learning을 위해 큰 dataset이 필요하지 않다. 50k 정도면 충분하고, 데이터 퀄리티가 더 중요하다.
Instruction Tuning Notion
 
 
Instruction Tuning Usages
 
 
 
 

Evaluation for Instruction tuning

Fine tuned LLM are more good at Positional Information processing such as Entity Tracking than base model. Position Transmitter or Value Fetcher mechanism using
Activation Patching
reproduced similar ability with Fine tuned model in base model.
Instruction tuned model’s
Attention head
are more attending on verbs to understand instruction.
 
 

 

Recommendations