Instruction Tuning

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

Self Instruct Tuning

Visual Instruction Tuning

Instruction hierarchy

Instruction Tuning Usages

Evaluation for Instruction tuning

Paper page - Instruction-Following Evaluation for Large Language Models

Join the discussion on this paper page

https://huggingface.co/papers/2311.07911

Imitation Models and the Open-Source LLM Revolution

Are proprietary LLMs like ChatGPT and GPT-4 actually easy to replicate?

https://cameronrwolfe.substack.com/p/imitation-models-and-the-open-source

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.

arxiv.org

https://arxiv.org/pdf/2402.14811

Instruction tuned model’s

Attention head are more attending on verbs to understand instruction.

aclanthology.org

https://aclanthology.org/2024.naacl-long.130.pdf

Instruction Tuning

Teach language models to follow instructions to solve a task

Evaluation for Instruction tuning

Recommendations