Fine-tuning and LoRA — teaching a model style without retraining it
LoRA fine-tunes a model by training a small number of added weights instead of all of them, cheap enough to run on one GPU — best for form and format, rarely the right tool for new facts.
Not yet scoped.
Why
The PoC fine-tunes a small model with LoRA on a style dataset and contrasts it with RAG, making the 'style vs facts' boundary the deliverable.
How it works
Not yet built.
Where it lands in Jayverse
Rabbit: use LoRA for a consistent agent voice, not for current facts. If Rabbit's agent needs a stable persona or support tone, that's LoRA's job; anything needing current market state or positions stays on RAG or a direct tool call.
Personas: keep a persona's voice separate from its factual claims. Treat any Personas-NFT chat behavior as the style-not-facts case this PoC tests, and source market facts from Verex's live data, never from fine-tuned weights.
Key expressions
Words and phrases from this page worth keeping, with the Korean meaning and the sentence they come from.
Expression
뜻 · 쓰이는 자리
cheap enough to
~할 만큼 저렴한 · "cheap enough to run on one GPU"
rarely the right tool for
~에는 좀처럼 적합한 도구가 아니다 · "rarely the right tool for new facts"
contrast X with Y
X를 Y와 대조하다 · "contrasts it with RAG"
the deliverable
결과물, 산출물 · "making the 'style vs facts' boundary the deliverable"
added weights
원 모델에 추가로 얹은 가중치 · "training a small number of added weights instead of all of them"
LoRA
저순위 적응(Low-Rank Adaptation) · 전체 가중치 대신 소수의 추가 가중치만 학습하는 파인튜닝 기법, 이 카드의 주제. "LoRA fine-tunes a model by training a small"
RAG
검색증강생성(Retrieval-Augmented Generation) · 스타일을 가르치는 LoRA와 대조되는, 사실을 보강하는 대안 기법. "and contrasts it with RAG"