I'm curious: What kind of projects can we develop using DeepSeek's open-source models? For example, what should we pay attention to when fine-tuning with custom data? What tools and methods are more efficient for a stable fine-tuning process? Additionally, what are the performance and cost advantages of running these models locally? What advice do you have for best practices that small teams or individual developers can use?
What can be done with DeepSeek's offerings?
👁️ 9 views💬 1 replies❤️ 0 likes
1 Replies
The possibilities with DeepSeek’s open-source models are almost limitless! I recently fine-tuned one for a client using our own data, and the results were quite satisfying. For example, we developed a custom chatbot to model customer support chats for an SME—cleaning the data and prompt engineering were crucial, but we kept getting positive feedback from the clients.
For small teams, the best approach is quickly fine-tuning using Hugging Face’s Transformers library and methods like LoRA (Low-Rank Adaptation). The advantage of running locally is clear: it saves costs, and with careful optimization, you can even get acceptable performance on a CPU without a GPU. I started with just a CPU for my first attempts, then saw a huge speed boost with a simple CUDA optimization—essential for tight budgets.