K8s-Operator (Fine-tuned LLaMA)

Author: Nir Adler Model Type: Fine-tuned LLaMA / Unsloth-based Model
Domain: Kubernetes (kubectl commands)
License: Apache 2.0 Repository: Model Page

🚀 Overview

This model is a fine-tuned LLaMA/Unsloth-based model designed to generate accurate and efficient kubectl commands for managing Kubernetes clusters. It understands Kubernetes concepts, CLI usage, and best practices, making it a valuable assistant for DevOps engineers, SREs, and platform teams.

It can help with:
✅ Constructing kubectl commands based on natural language queries.
✅ Explaining Kubernetes commands and best practices.
✅ Providing structured responses with safe and efficient command execution.

📖 Model Details

Base Model: llama-3.2-3b-instruct-unsloth-bnb-4bit
Fine-tuned on:
- ComponentSoft/k8s-kubectl (General k8s commands)
- ComponentSoft/k8s-kubectl-35k (Expanded dataset)
- ComponentSoft/k8s-kubectl-cot-20k (Chain of Thought explanations)
Training Framework: Unsloth (optimized for efficient training)
Format: ShareGPT-style chat template
Dataset Size: ~55K Kubernetes-related command pairs

🔧 Usage

1️⃣ Load the Model from Hugging Face

from transformers import AutoModelForCausalLM, AutoTokenizer

model_name = "niradler/k8s_operator"
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForCausalLM.from_pretrained(model_name)

3️⃣ Example Output

## **🔹 Instruction:**  
Retrieve logs from a running pod named `web-server`.  

## **🔹 Recommended `kubectl` Command:**  
```sh
kubectl logs web-server

📌 Limitations & Considerations

🔹 The model is not aware of Kubernetes cluster states, so verify command outputs.
🔹 It may generate destructive commands—always review before running.
🔹 Fine-tuning with specific Kubernetes versions might be necessary for up-to-date command compatibility.

🔗 Resources & Links

🤗 Hugging Face Model Page: Hugging Face Link
📜 Training Dataset:
📘 Official Kubernetes Docs: https://kubernetes.io/docs/
🛠️ Fine-tuning with Unsloth: https://unsloth.ai/