Let’s Talk Tokens: AMA on Reinforcement Fine-Tuning (RFT), GRPO, and AI Rewards

🔔 SUBSCRIBE for the latest on LLM fine-tuning, AI scaling, and reinforcement learning hacks! 👉https://www.youtube.com/@Predibase Reinforcement Fine-Tuning (RFT) is no longer theoretical—it’s powering real-world GenAI systems today. In this live AMA, Predibase’s AI experts answer audience questions on: ✅ When to use RFT over SFT ✅ How GRPO (Group Relative Policy Optimization) works ✅ Designing robust reward functions (and avoiding reward hacking) ✅ How much data you really need for text-to-SQL, codegen, and logic tasks ✅ Future of RFT in agentic workflows and enterprise GenAI systems This is a deep dive into best practices, landmines, and emerging strategies from the engineers behind the DeepLearning.AI course on RFT + GRPO and the Predibase RFT platform. 🧠 Speakers: • Travis Addair – CTO & Co-founder, Predibase • Arnav Garg – ML Engineering Lead, Predibase • Ajinkya Tejankar – Senior Research Engineer, Predibase 🔗 Try Predibase’s RFT Platform: https://predibase.com/free-trial 👉 Schedule a live demo: https://pbase.ai/41FZKfy 00:00 — Intro & Welcome 02:00 — What Is RFT and Why It Matters Now 07:00 — GRPO vs PPO vs RLHF Explained 11:00 — Use Cases Best Suited for RFT 15:30 — When to Use RFT vs SFT 20:00 — How Much Data Do You Need for RFT? 25:00 — GRPO Hyperparameters & Prompt Design 30:00 — Reward Function Design & Best Practices 37:00 — Common RFT Pitfalls and Reward Hacking 43:00 — Handling Subjective Evaluations 48:00 — Future of RFT: Agentic AI & Beyond 52:00 — Final Thoughts + AMA Wrap-up #reinforcementlearning #rft #grpo #aitraining #genai #llms #finetuning #machinelearning #agenticai #RLHF #predibase #AIinProduction #llminference