# Practical Guide to Function Calling Fine-Tuning for Turkish Large Language Models: Optimization of the Kizagan Model Using Unsloth and QLoRA

> A complete Jupyter Notebook tutorial demonstrating how to add function calling capabilities to a Turkish reasoning model using Unsloth and QLoRA technologies on an H100 GPU, covering the entire workflow from environment setup to model deployment.

- 板块: [Openclaw Llm](https://www.zingnex.cn/en/forum/board/openclaw-llm)
- 发布时间: 2026-05-26T08:46:25.000Z
- 最近活动: 2026-05-26T08:52:09.912Z
- 热度: 163.9
- 关键词: Turkish LLM, function calling, Unsloth, QLoRA, fine-tuning, H100, Flash Attention, LoRA, catastrophic forgetting, reasoning model
- 页面链接: https://www.zingnex.cn/en/forum/thread/unslothqlorakizagan
- Canonical: https://www.zingnex.cn/forum/thread/unslothqlorakizagan
- Markdown 来源: floors_fallback

---

## Introduction / Main Floor: Practical Guide to Function Calling Fine-Tuning for Turkish Large Language Models: Optimization of the Kizagan Model Using Unsloth and QLoRA

A complete Jupyter Notebook tutorial demonstrating how to add function calling capabilities to a Turkish reasoning model using Unsloth and QLoRA technologies on an H100 GPU, covering the entire workflow from environment setup to model deployment.

## Original Author and Source

- **Original Author/Maintainer**: Tuguberk
- **Source Platform**: GitHub
- **Original Title**: Kizagan-E4B-FunctionCalling-TR
- **Original Link**: https://github.com/Tuguberk/Kizagan-E4B-FunctionCalling-TR
- **Publication Date**: 2026-05-26

---

## Project Overview

This is a complete fine-tuning tutorial for Turkish large language models, aiming to add function calling capabilities to a base reasoning model. The project uses the Unsloth framework combined with QLoRA technology, optimized specifically for the NVIDIA H100 GPU, fully leveraging the bfloat16 and Flash Attention 2 features of the Hopper architecture.

The base model used is `AlicanKiraz0/Kizagan-E4B-Turkish-Reasoning-Model`, a Turkish model focused on reasoning capabilities. The fine-tuned model will have both function calling and general Turkish dialogue capabilities.

---

## Unsloth Framework

Unsloth is one of the most popular optimization frameworks in the field of large model fine-tuning currently. Its core advantages include:

- **Automatic Flash Attention 2 Activation**: Automatically enabled on supported hardware, significantly improving training speed
- **4-bit NF4 Quantization**: Greatly reduces memory usage, making large model fine-tuning possible on consumer-grade hardware
- **bfloat16 Optimization**: Natively compatible with H100, more stable than fp16
- **Gradient Checkpoint Integration**: Unsloth provides a specially optimized version, balancing memory and speed

## QLoRA (Quantized Low-Rank Adaptation)

QLoRA is the mainstream method for Parameter-Efficient Fine-Tuning (PEFT) currently. Its working principle is:

1. **Freeze Base Model Parameters**: Keep pre-trained weights unchanged to avoid catastrophic forgetting
2. **Inject Trainable Low-Rank Matrices**: Add small adapters to key layers (q_proj, k_proj, v_proj, o_proj, etc.)
3. **4-bit Quantized Storage**: The base model is stored in NF4 format, greatly reducing memory requirements
4. **Dequantization During Training**: Dynamically restore to bf16 during computation to ensure accuracy

The LoRA parameters configured in this project are r=32 and alpha=32, which are relatively high configurations and can learn more complex patterns.

## Flash Attention 2

Flash Attention is an IO-aware exact attention algorithm. Through block computation and recomputation strategies, it reduces the memory complexity of standard attention from O(N²) to O(N). On the H100, this can bring a 2-3x improvement in training speed.

---

## Dataset Strategy: Preventing Catastrophic Forgetting

The project uses a mixed dataset strategy, which is a key consideration in function calling fine-tuning:

## Dataset Composition

| Dataset | Ratio | Purpose |
|---------|-------|---------|
| Turkish Hermes Function Calling | 85% | Function calling training data, containing 144K Turkish prompt/completion pairs |
| Turkish LLM v10 Training |15% | General Turkish dialogue data to maintain basic language capabilities |