Post-Training AI

Pre-trained language models can generate text, but they don't reliably follow instructions, reason through tasks, or use tools like agents. Post-training adds those capabilities. In Post-Training AI, Ben Burtenshaw, a machine learning engineer at Hugging Face, shows practitioners how to turn a base model into a system for production, with the control that comes from owning the model weights.

You'll build post-training systems from the ground up, starting with supervised fine-tuning and moving into preference learning, and reinforcement learning with verifiable rewards. Using tools from the Hugging Face ecosystem, you'll go from experiments on a single GPU to pipelines that scale, often at a fraction of the cost of relying on frontier APIs. By the end, you'll know how to shape model behavior for your domain, improve reliability, and choose the right techniques for your use case.

• Build supervised fine-tuning and reinforcement learning projects from first principles
• Post-train open-weight models with TRL
• Design post-training pipelines that scale with PEFT and distributed systems
• Build agentic execution environments that train models to interact with the real world
• Evaluate models using benchmarks, rubrics, and failure-mode analysis