# Llama 3.1 / 3.2 / 3.3 8B — RTX 4090'ın İş Atı: GQA + 128K Context + Türkçe Reçete

> Source: https://sukruyusufkaya.com/learn/fine-tuning-cookbook/ftc-llama-3x-8b-rtx4090-recipe
> Updated: 2026-08-15T06:47:59.136Z
> Category: Fine-Tuning Cookbook (Model-by-Model)
> Module: Part III — Small Open Models (1B–8B)
**TLDR:** Llama 3.1/3.2/3.3 8B-Instruct'ın anatomisi: 32 layer × 4096 hidden, GQA (8 KV-head), RoPE θ=500K, SwiGLU, RMSNorm, 128K context (YaRN-extended). RTX 4090'da QLoRA NF4 + Unsloth ile 50K Türkçe Alpaca üzerinde 1 epoch ~50 dakika. TR-MMLU baseline 32.4 → fine-tune 39.8 (+%23). Full reçete: dataset format, hyperparameter tablosu, sweep aralıkları, sample inference, eval pipeline.