Uploaded May 2025 | Updated September 2026, 2 weeks ago
Learn more: bit.ly/43p1WIa
DeepSeek has put reinforcement learning at the top of the minds of developers, machine learning engineers, and data-driven professionals in the AI space.
That’s why we’re happy to launch a new short course: Reinforcement Fine-Tuning LLMs with GRPO, built in collaboration with @Predibase and taught by Travis Addair, its Co-Founder and CTO, and Arnav Garg, its Senior Machine Learning Engineer.
Many LLM applications rely on reasoning, whether in solving math problems, generating code, or completing multi-step tasks. But fine-tuning models for reasoning is often constrained by the availability of high-quality labeled examples.
This course introduces a different approach: Reinforcement Fine-Tuning (RFT) using Group Relative Policy Optimization (GRPO). GRPO is a scalable reinforcement learning algorithm that lets you train models using reward functions instead of human-labeled data or preference scores.
You’ll learn:
- When reinforcement fine-tuning is a better fit than supervised fine-tuning
- How to build and use programmable reward functions in GRPO
- How to guide model behavior on structured tasks like the Wordle game
- How to evaluate subjective outputs, like summaries, using LLMs as judges
- How to avoid reward hacking by combining reward and penalty signals
- How to implement GRPO loss: token ratios, clipping, advantages, and KL divergence
- How to run RFT jobs using Predibase’s training platform
By the end of the course, you’ll know how to fine-tune LLMs for complex reasoning tasks without needing large datasets or manual preference data.
Enroll now: bit.ly/43p1WIa
Learn more: bit.ly/43p1WIa
DeepSeek has put reinforcement learning at the top of the minds of developers, machine learning engineers, and data-driven professionals in the AI space.
That’s why we’re happy to launch a new short course: Reinforcement Fine-Tuning LLMs with GRPO, built in collaboration with @Predibase and taught by Travis Addair, its Co-Founder and CTO, and Arnav Garg, its Senior Machine Learning Engineer.
Many LLM applications rely on reasoning, whether in solving math problems, generating code, or completing multi-step tasks. But fine-tuning models for reasoning is often constrained by the availability of high-quality labeled examples.
This course introduces a different approach: Reinforcement Fine-Tuning (RFT) using Group Relative Policy Optimization (GRPO). GRPO is a scalable reinforcement learning algorithm that lets you train models using reward functions instead of human-labeled data or preference scores.
You’ll learn:
- When reinforcement fine-tuning is a better fit than supervised fine-tuning
- How to build and use programmable reward functions in GRPO
- How to guide model behavior on structured tasks like the Wordle game
- How to evaluate subjective outputs, like summaries, using LLMs as judges
- How to avoid reward hacking by combining reward and penalty signals
- How to implement GRPO loss: token ratios, clipping, advantages, and KL divergence
- How to run RFT jobs using Predibase’s training platform
By the end of the course, you’ll know how to fine-tune LLMs for complex reasoning tasks without needing large datasets or manual preference data.
Enroll now: bit.ly/43p1WIa
