Hello, I've just started working with DeepSeek-style models. Many things confuse me, from their logic to how we prepare the data. What are the general approaches to prompt engineering and in-context learning, specifically? How can I get the best performance out of the system? If anyone has experience, can you share your methods? I'm looking for a general overview, not specific experiences.
How should one approach DeepSeek? General methods
👁️ 4 views💬 1 replies❤️ 0 likes
1 Replies
When starting with these models, the most important thing is to realize that they are actually just "tools." At first, I used to believe they were magical sorcerers, but later I saw that *how* you present things makes a huge difference. What we call prompt engineering is essentially guiding them: set clear goals, avoid unnecessary details, and—especially—clearly specify how you want the outputs. For example, instead of saying, "Summarize this text in 5 sentences," giving detailed instructions like "5 bullet points, each no more than 10 words, in an academic tone" makes the results much more useful.
As for in-context learning, you need to observe how the model responds based on how you present the data. I tested this on small projects: I gave short examples and then checked how similar the model's responses were. For instance, if you're generating SEO-friendly meta descriptions for a website, a tried-and-tested method is to preload it with 3-4 example sentences to get the best results. Additionally, sending the same prompt at different times and comparing the outputs is useful for testing the model's consistency. Generally, it takes some time to discover the system's limits, but patience is key.