# SFT on Reasoning Traces: Llama-8B + R1-Distilled Traces (8K → 32K Context)

> Source: https://sukruyusufkaya.com/learn/fine-tuning-cookbook/ftc-sft-reasoning-traces
> Updated: 2026-08-10T22:57:40.875Z
> Category: Fine-Tuning Cookbook (Model-by-Model)
> Module: Part XII — Reasoning Model FT (R1-style)
**TLDR:** Reasoning trace dataset hazırsa SFT teknik olarak basit ama detay önemli: \<think\> token vocab'a ekleme, embedding init, context length 32K (R1 traces 5-15K token), loss masking (think tokens loss'a girer veya girmez?), epoch count. RTX 4090 + Llama 3.1 8B + 1000 R1 trace 1 epoch ~50 dakika.