# GRPO (Group Relative Policy Optimization): DeepSeek-R1'in Verifiable Reward Reçetesi

> Source: https://sukruyusufkaya.com/learn/fine-tuning-cookbook/ftc-grpo-deepseek-r1-verifiable-reward
> Updated: 2026-08-22T02:56:46.547Z
> Category: Fine-Tuning Cookbook (Model-by-Model)
> Module: Part XI — Alignment & Preference Optimization
**TLDR:** GRPO (DeepSeek 2024) — PPO'nun simplified varyantı. Critic/value head yok. Bir batch'te G adet farklı response sample et, group içinde **göreli reward**'ları normalize et. Verifiable rewards (math correctness, code execution) ile reasoning RL'i mümkün kıl. RTX 4090'da Qwen-7B + GRPO + GSM8K accuracy +%5-8.