The Craft of Post-Training - Chris Von Csefalvay - diskuze

If you're a practitioner who has watched a promising AI demo fail to survive contact with production, where prompting hits its ceiling, retrieval isn't enough, and the model still can't be trusted with your domain, post-training is what you've been missing. The Craft of Post-Training is a practical guide to turning foundation models into production-ready systems - reshaping behavior, aligning to your values, and deploying with confidence. Each technique is taught concept-first, then implementation-through-code, so you understand not just what to run, but what you're actually changing inside the model. You'll leave with the skills to: Fine-tune models on curated datasets using supervised fine-tuning, LoRA, and QLoRA without destroying the base model's general capabilities; Apply reinforcement learning from human feedback and modern preference optimization methods, including GRPO, ORPO, and beyond, to shape model behavior; Evaluate models rigorously: design benchmarks, detect regression, and measure quality claims that survive scrutiny; Adapt models to specialized domains, from clinical language to legal text, turning general capability into a defensible competitive advantage; Train agentic models that take sequences of actions reliably, not just models that talk about taking actions; Quantize and compress fine-tuned models for deployment without sacrificing the gains you trained for. Post-training is where models stop being impressive and start being useful. This book teaches
více informací