Fine-tuned Gemma 3 12B model specialized for answering Kubernetes questions in Turkish, quantized to 4bit format for efficient inference with reduced memory footprint.

Model Description

This repository contains a 4bit quantized version of the Kubernetes AI model, optimized for running on consumer hardware with reduced VRAM/RAM requirements. The model uses BitsAndBytes quantization with safetensors format for fast loading and efficient inference.

Primary Purpose: Answer Kubernetes-related questions in Turkish language with minimal hardware requirements.

Model Specifications

Specification	Details
Format	Safetensors (4bit quantized)
Base Model	unsloth/gemma-3-12b-it-qat-bnb-4bit
Quantization	4bit (BitsAndBytes)
Model Size	~7.2 GB
Memory Usage	~8-10 GB VRAM/RAM
Precision	4bit weights, FP16 compute

Quick Start

Installation

bash

# Install required packages
pip install torch transformers accelerate bitsandbytes safetensors

Basic Usage

python

from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

# Load model and tokenizer
model_name = "aciklab/kubernetes-ai-4bit"
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForCausalLM.from_pretrained(
    model_name,
    device_map="auto",
    trust_remote_code=True
)

# Prepare input
prompt = "Kubernetes'te 3 replikaya sahip bir deployment nasıl oluştururum?"

# Format with chat template
messages = [
    {"role": "system", "content": "Sen Kubernetes konusunda uzmanlaşmış bir yapay zeka asistanısın. Kubernetes ile ilgili soruları Türkçe olarak yanıtlıyorsun."},
    {"role": "user", "content": prompt}
]

input_text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tokenizer(input_text, return_tensors="pt").to(model.device)

# Generate response
outputs = model.generate(
    **inputs,
    max_new_tokens=512,
    temperature=1.0,
    top_p=0.95,
    top_k=64,
    repetition_penalty=1.05,
    do_sample=True
)

response = tokenizer.decode(outputs[0], skip_special_tokens=True)
print(response)

Training Details

This model is based on the aciklab/kubernetes-ai LoRA adapters:

Base Model: unsloth/gemma-3-12b-it-qat-bnb-4bit
Training Method: LoRA (Low-Rank Adaptation)
LoRA Rank: 8
Target Modules: q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj
Training Dataset: ~157,210 examples from Kubernetes docs, Stack Overflow, and DevOps datasets
Training Time: 28 hours on NVIDIA RTX 5070 12GB
Max Sequence Length: 1024 tokens

Training Dataset Summary

Dataset Category	Count	Description
Kubernetes Official Docs	8,910	Concepts, kubectl, setup, tasks, tutorials
Stack Overflow	52,000	Kubernetes Q&A from community
DevOps Datasets	62,500	General DevOps and Kubernetes content
Configurations & CLI	36,800	Kubernetes configs, kubectl examples, operators
Total	~157,210	Comprehensive Kubernetes knowledge base

Quantization Details

This model uses 4bit quantization with BitsAndBytes for optimal memory efficiency:

Source: Merged LoRA adapters with base model
Quantization Method: BitsAndBytes 4bit (NF4)
Compute Precision: FP16
Format: Safetensors (fast loading)
Memory Footprint: ~7.2 GB on disk, ~8-10 GB in memory

Advantages of 4bit Format

Efficient Memory Usage: Runs on GPUs with 8GB+ VRAM
Fast Loading: Safetensors format loads quickly
Good Quality: Minimal accuracy loss compared to full precision
Framework Support: Compatible with Transformers, vLLM, Text Generation Inference
Flexible Deployment: Can run on CPU with acceptable speed

Hardware Requirements

Minimum (GPU)

GPU: 8GB VRAM
RAM: 8GB system memory
Storage: 10GB free space

Limitations

Language: Optimized primarily for Turkish and English.
Domain: Specialized for Kubernetes; may not perform well on general topics
Quantization: 4bit quantization may occasionally affect response quality on complex queries

License

This model is released under the MIT License. Free to use in commercial and open-source projects.

Citation

If you use this model in your research or applications, please cite:

bibtex

@misc{kubernetes-ai-4bit,
  author = {HAVELSAN/Açıklab},
  title = {Kubernetes AI - 4bit Safetensors},
  year = {2025},
  publisher = {HuggingFace},
  howpublished = {\url{https://huggingface.co/aciklab/kubernetes-ai-4bit}}
}

Contact

Produced by: HAVELSAN/Açıklab

For questions, feedback, or issues, please open an issue on the model repository or contact us through HuggingFace.

Related Models

aciklab/kubernetes-ai - Original LoRA adapters
aciklab/kubernetes-ai-GGUF - GGUF quantized versions for llama.cpp

Note: This is a 4bit quantized model ready for immediate use with the Transformers library. No additional model merging or quantization required.

aciklab/kubernetes-ai-4bit

Kubernetes AI - 4bit Safetensors