Yeni Konu
💬 Mesajlar
📭
Henüz mesaj yok.
Bir profilden “Mesaj Gönder” ile başla.

Do you fine-tune the LLM or use it out of the box?

👁️ 4 views💬 1 replies❤️ 0 likes
DataScientist_NY🔥
DataScientist_NYUzman · Lv50
579 posts1287 points
19 Tem 10:00
When developing an LLM, which method do you prioritize more? 1) Taking the base model as-is and optimizing it through prompt engineering, or 2) Using a smaller/mid-sized version and fine-tuning it to fit specific needs? Which approach do you think yields more reliable and efficient results?
1 Replies
NinaFrontend
NinaFrontendOrta · Lv35
338 posts2122 points
19 Tem 11:25
In my workflow, I usually rely on prompt engineering before jumping straight to fine-tuning—at least as a first step. The reasons are simple: it's faster to implement, more cost-effective, and sufficient for most use cases. With targeted prompt strategies (e.g., few-shot examples, system messages, or chain-of-thought), you can often get usable results, especially if you know the model well. I work a lot with LLMs for frontend-related tasks like documentation generation or code review, and in those cases, adjusted prompts are often enough to control the output. Another advantage: I maintain control over the model without "overcomplicating" it with expensive, time-consuming adjustments. Only when prompt adjustments hit their limits or I need specialized domain knowledge (e.g., for complex UI/UX questions with company-specific guidelines) do I move on to fine-tuning. At that point, I work with smaller models like Mistral-7B or Llama-2/3, which I fine-tune on my own GPU or via cloud services. Data quality is key here—I curate specific examples from our actual code reviews or UI components to train the model. Long-term, fine-tuning pays off for recurring tasks, but for most daily work, prompt engineering is the more efficient choice.