aciklab/kubernetes-ai-lora

Kubernetes AI - Gemma 3 12B LoRA Adapters

Fine-tuned Gemma 3 12B model specialized for answering Kubernetes questions in Turkish.

Model Description

This model consists of LoRA adapters fine-tuned on unsloth/gemma-3-12b-it-qat-bnb-4bit using a comprehensive dataset of Kubernetes documentation, Stack Overflow questions, and DevOps scenarios.

Primary Purpose: Answer Kubernetes-related questions in Turkish language.

Use Cases

• Kubernetes cluster management and troubleshooting
• YAML configuration generation and validation
• kubectl command assistance
• Debugging pod, service, and deployment issues
• Kubernetes best practices and concepts
• DevOps workflow optimization
• Turkish language Kubernetes Q&A

Quick Start

Loading the Model

from unsloth import FastLanguageModel
from peft import PeftModel
import torch

# Load base Gemma 3 12B model
model, tokenizer = FastLanguageModel.from_pretrained(
    model_name="unsloth/gemma-3-12b-it-qat-bnb-4bit",
    max_seq_length=2048,
    dtype=None,
    load_in_4bit=True,  # Use 4-bit quantization to fit in GPU memory
)

# Load Kubernetes AI LoRA adapters
model = PeftModel.from_pretrained(
    model,
    "aciklab/kubernetes-ai"
)

# Enable inference mode
FastLanguageModel.for_inference(model)

# Example usage (Turkish question)
messages = [
    {"role": "user", "content": "Kubernetes'te 3 replikaya sahip bir deployment nasıl oluştururum?"}
]

inputs = tokenizer.apply_chat_template(
    messages,
    tokenize=True,
    add_generation_prompt=True,
    return_tensors="pt"
).to("cuda")

outputs = model.generate(
    input_ids=inputs,
    max_new_tokens=512,
    temperature=0.7,
    top_p=0.9,
    do_sample=True
)

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

Example Questions

Turkish Examples

# Deployment creation
"Node.js uygulaması için 3 replika, sağlık kontrolleri ve kaynak limitleri olan bir Kubernetes deployment oluştur."

# Troubleshooting
"Pod'um CrashLoopBackOff durumunda. Yaygın nedenleri nelerdir ve nasıl debug ederim?"

# kubectl commands
"Production namespace'indeki çalışmayan tüm pod'ları gösteren kubectl komutunu yaz."

# Best practices
"Kubernetes'te container güvenliği için en iyi uygulamalar nelerdir?"

# Service creation
"LoadBalancer tipinde bir Kubernetes servisi nasıl yapılandırılır?"

English Examples

"How do I create a Kubernetes deployment with 3 replicas?"
"What are the common causes of CrashLoopBackOff?"
"Show me kubectl command to get all pods in production namespace."

Training Dataset

The model was trained on ~157,000 examples from multiple high-quality Kubernetes and DevOps datasets:

Dataset	Count	Description
Kubernetes Official Documentation
- Concepts	2,700	Core Kubernetes concepts
- Kubectl Reference	600	kubectl command documentation
- Setup Guides	430	Installation and setup
- Tasks	4,300	Practical task guides
- Tutorials	880	Step-by-step tutorials
Stack Overflow
mcipriano/stackoverflow-kubernetes-questions	30,000	Kubernetes Q&A
peterpanpan/stackoverflow-kubernetes-questions	22,000	Additional Kubernetes Q&A
DevOps Datasets
Szaid3680/Devops	42,000	General DevOps content
ahmedgongi/Devops_LLM	20,500	Kubernetes-filtered DevOps (from 140k)
Configuration & Operations
HelloBoieeee/kubernetes_config	10,000	Kubernetes configurations
sidddddddddddd/kubernetes-with-ood	6,000	Kubernetes scenarios (incl. Turkish translations)
dereklck/kubernetes_cli_dataset_20k	19,000	kubectl CLI examples
dereklck/kubernetes_operator_3b_1.5k	1,800	Kubernetes operator patterns

Total Training Examples: ~157,210

Training Details

• Base Model: unsloth/gemma-3-12b-it-qat-bnb-4bit
• Method: LoRA (Low-Rank Adaptation)
• Framework: Unsloth
• LoRA Rank: 8
• Target Modules: q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj
• Training Checkpoint: checkpoint-8175
• Max Sequence Length: 1024 tokens
• Training Time: 28 hours
• Hardware: NVIDIA GeForce RTX 5070 12GB

Hardware Requirements

• Minimum VRAM: 12GB (with 4-bit quantization)
• Recommended VRAM: 24GB (for faster inference)
• CPU RAM: 32GB+
• Training Hardware: RTX 5070 12GB

Limitations

• May not have information on very recent Kubernetes features released after training
• Primarily trained for Turkish language responses, though it can handle English queries
• Best suited for technical Kubernetes questions; general conversation capabilities can be limited

Performance Notes

• Trained on RTX 5070 12GB in 28 hours
• Works with 12GB VRAM using 4-bit quantization
• Fast startup by loading only adapters without full model reload

License

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

Acknowledgments

• Google and Unsloth team for the Gemma 3 base model
• Unsloth team for the efficient training framework
• All dataset contributors
• Kubernetes community for comprehensive documentation
• NVIDIA for RTX 5070 enabling 28-hour training

Contact

For questions or feedback, please open an issue on the model repository.

---

Note: This is a LoRA adapter, not a full model. You must load it on top of unsloth/gemma-3-12b-it-qat-bnb-4bit to use it.

Related Links

• Unsloth Documentation
• Gemma Model Card
• PEFT Documentation
• Kubernetes Documentation

Citations

Datasets

@misc{stackoverflow-kubernetes-mcipriano,
  author = {mcipriano},
  title = {Stack Overflow Kubernetes Questions},
  year = {2024},
  publisher = {HuggingFace},
  url = {https://huggingface.co/datasets/mcipriano/stackoverflow-kubernetes-questions}
}

@misc{devops-szaid,
  author = {Szaid3680},
  title = {DevOps Dataset},
  year = {2024},
  publisher = {HuggingFace},
  url = {https://huggingface.co/datasets/Szaid3680/Devops}
}

@misc{devops-llm-ahmed,
  author = {ahmedgongi},
  title = {DevOps LLM Dataset},
  year = {2024},
  publisher = {HuggingFace},
  url = {https://huggingface.co/datasets/ahmedgongi/Devops_LLM}
}

@misc{kubernetes-config-hello,
  author = {HelloBoieeee},
  title = {Kubernetes Config Dataset},
  year = {2024},
  publisher = {HuggingFace},
  url = {https://huggingface.co/datasets/HelloBoieeee/kubernetes_config}
}

@misc{kubernetes-ood-sidddddddddddd,
  author = {sidddddddddddd},
  title = {Kubernetes with OOD Dataset},
  year = {2024},
  publisher = {HuggingFace},
  url = {https://huggingface.co/datasets/sidddddddddddd/kubernetes-with-ood}
}

@misc{stackoverflow-kubernetes-peter,
  author = {peterpanpan},
  title = {Stack Overflow Kubernetes Questions},
  year = {2024},
  publisher = {HuggingFace},
  url = {https://huggingface.co/datasets/peterpanpan/stackoverflow-kubernetes-questions}
}

@misc{kubernetes-operator-derek,
  author = {dereklck},
  title = {Kubernetes Operator Dataset},
  year = {2024},
  publisher = {HuggingFace},
  url = {https://huggingface.co/datasets/dereklck/kubernetes_operator_3b_1.5k}
}

@misc{kubernetes-cli-derek,
  author = {dereklck},
  title = {Kubernetes CLI Dataset},
  year = {2024},
  publisher = {HuggingFace},
  url = {https://huggingface.co/datasets/dereklck/kubernetes_cli_dataset_20k}
}

Model

@misc{kubernetes-ai,
  author = {aciklab},
  title = {Kubernetes AI Turkish - Gemma 3 12B LoRA Adapters},
  year = {2025},
  publisher = {HuggingFace},
  url = {https://huggingface.co/aciklab/kubernetes-ai},
  note = {Trained on RTX 5070 12GB in 28 hours}
}