Observing FineTuning and RAG being used interchangeably

Haebom

Nov 27, 20232y ago

Recently, I've noticed that two terms are often used interchangeably when reading posts on Facebook or LinkedIn. As their names suggest, they're clearly different methods. Both are used to improve model performance and adjust models for specific tasks or contexts. That's probably why they're sometimes mixed up. Here are the differences between the two.

Fine-tuning

•

Rather than adding new information to an existing model, the goal is to modify the model’s behavior.

•

Typical cases for fine-tuning include ensuring the model gives a certain response, or producing complex outputs that require a sequence of prompts.

•

Fine-tuning can also help reduce the number of tokens needed or shorten request latency in many scenarios.

Retrieval-augmented Generation (RAG)

•

If you want to add new knowledge to your model, you should use RAG.

•

RAG doesn’t directly add new knowledge, but supplies relevant information to the prompt in the request. The model then uses this to answer your question.

•

In experiments with real datasets comparing a fine-tuned model and a base model combined with RAG, the fine-tuned model had 0% accuracy, while the model with RAG reached 95%. This clearly demonstrates that fine-tuning does not add new knowledge to the model.

Differences:

•

Purpose and operation: Fine-tuning focuses on adapting or specializing how the base model behaves, while RAG draws on external data sources to enrich the model’s output.

•

Knowledge integration: While RAG seems to add new knowledge to the model, it actually uses search results rather than changing the model directly. Fine-tuning, on the other hand, alters the model itself.

These two approaches are often confused because both are used to improve model outputs and may give similar results in certain cases. However, their core goals and how they work are quite different. RAG generates responses using external resources, while fine-tuning tweaks the model’s built-in knowledge and learning process.

Recommended reading

Skelter Labs Blog - RAG vs. 파인튜닝 :: 기업용 맞춤 LLM을 위한 선택 가이드

기업이 원하는 맞춤 LLM 모델을 만드는 방법은 여러가지가 있습니다. 그 중 RAG 와 파인튜닝의 특장점과 기업이 도입할 땐 어떤 방법이 더 효율적일지 상세히 알아봅니다.

skelterlabs.com

SELF-RAG: 언어 모델의 한계를 넘어서 - 해봄의 잡동사니

언어 모델이 계속 발전하면서 다양한 실험들이 계속 되고 있습니다. 그 중에서도 Retrieval Augmented Generation(RAG)은 언어 모델이 다양한 문맥을 이해하고 생성할 수 있게 하는 기술로 주목받고 있습니다. 하지만 RAG에도 한계가 있습니다. 이러한 한계를 극복하기 위한 새로운 연구가 등장했는데, 바로 워싱턴 대학교와 IBM 연구팀이 ...

slashpage.com

Subscribe to 'haebom'

📚 Welcome to Haebom's archives.
---
I post articles related to IT 💻, economy 💰, and humanities 🎭.
If you are curious about my thoughts, perspectives or interests, please subscribe.
haebom@kakao.com