دانلود کتاب دستور العمل های مدل زبانی بزرگ – راهنمای عملی برای تنظیم دقیق، بهینه سازی، استقرار و برنامه های کاربردی در دنیای واقعی

There is no shortage of LLM books today. Some go deep into the mathematics, leaving you wondering how any of it translates into a working system. Others hand you code snippets and expect you to figure out the rest. We wrote LLM Recipes because we wanted a book that does both: one that explains ideas clearly enough to build real intuition and backs every concept up with code you can actually run. What sets this book apart starts with how it teaches. Rather than jumping straight to code, every major concept is introduced through carefully designed figures, architecture diagrams, and visual walkthroughs. We believe a good diagram can make something click in seconds that paragraphs of text alone struggle to convey. When we explain attention mechanisms, LoRA adapters, or a RAG pipeline, the figure comes first. The code follows once you already have a mental picture of what is happening. Figures like the LLM Data Preparation Workflow (Figure 4-1), the LoRA Weight Decomposition (Figure 8-2), and the RAG Pipeline Architecture (Figure 13-3) are not decorations; they are reference anchors you will return to throughout the book and in your own work. The second distinction is scope. Most practical LLM guides cover a slice of the problem: a fine-tuning book here, a deployment manual there, a chapter on quantization somewhere else. LLM Recipes covers the entire post-training journey in one volume, from prompt engineering and instruction tuning through parameter-efficient fine-tuning, quantization, production deployment, evaluation, RAG, vision language models, and agentic workflows. We wrote it this way because these stages are deeply linked in practice. Your quantization choices affect serving latency. Your instruction format shapes your prompting surface. Your evaluation discipline determines every decision upstream. Treating them in isolation misses the point. We should also be upfront about something practical: dedicated GPUs are still out of reach for many practitioners, students, and independent researchers. A cookbook that requires an A100 cluster is not practical. That is why we carefully selected every model, configuration, and optimization strategy in this book to run on free-tier Google Colab GPUs. Every recipe has been tested in that environment. Where a technique benefits from more compute, we say so, but the default path always works on accessible hardware, whether you are in a university lab, on a coffee shop laptop, or at a personal machine with a single consumer GPU.