lalatendu/phi3-sysadmin-lalatendu

phi3-sysadmin-lalatendu

phi3-sysadmin-lalatendu is a domain-specialized model based on microsoft/Phi-3-mini-4k-instruct, fine-tuned using QLoRA (SFT with LoRA) via Unsloth for Linux system administration and DevOps tasks. This repository provides the GGUF (Q4_K_M) quantized model ready for local inference via Ollama.

• —Developed by: Lalatendu Keshari Swain
• —Model type: Causal Language Model (GGUF quantized)
• —Language(s): English
• —License: MIT
• —Base model: microsoft/Phi-3-mini-4k-instruct (3.8B parameters)
• —Fine-tuning method: QLoRA (Supervised Fine-Tuning with LoRA)
• —Quantization: q4km (4-bit, ~2.3 GB)

Disclaimer: This model is provided for educational and productivity purposes only. We take no responsibility for the accuracy or completeness of the outputs. Commands and configurations suggested by this model should always be verified by a qualified system administrator before being applied to any production system. Please use it at your own risk.

Model Sources

• —LoRA Adapter: lalatendu/phi3-sysadmin-lora
• —GitHub: github.com/lalatenduswain
• —Blog: blog.lalatendu.info

Training Process

This model was trained using a single-stage SFT process:

Step 1: SFT (Supervised Fine-Tuning)

• —Dataset: 1,026 curated sysadmin and DevOps Q&A examples in ChatML JSONL format
• —Format: system / user / assistant turns
• —Topics: Linux administration, AWS, Docker, Kubernetes, Terraform, Ansible, Nginx, databases, networking, security, monitoring, backup
• —Objective: To specialize the Phi-3 Mini model in answering practical server management and troubleshooting questions accurately and concisely.

Training Hyperparameters

GGUF Export

• —Quantization method: q4km via llama.cpp
• —File size: ~2.3 GB
• —Export tool: Unsloth's built-in GGUF exporter

How to Get Started

Option 1: Ollama (Recommended)

# 1. Install Ollama (if not already installed)
curl -fsSL https://ollama.com/install.sh | sh

# 2. Download phi3-sysadmin-Q4_K_M.gguf and Modelfile from this repo

# 3. Create the model
ollama create phi3-sysadmin -f Modelfile

# 4. Run interactively
ollama run phi3-sysadmin

Example queries:

ollama run phi3-sysadmin "How do I find what's consuming disk space?"
ollama run phi3-sysadmin "How do I set up Nginx reverse proxy with SSL?"
ollama run phi3-sysadmin "How do I troubleshoot high CPU usage?"
ollama run phi3-sysadmin "How do I create a Kubernetes deployment?"

API usage:

curl http://localhost:11434/api/generate -d '{
  "model": "phi3-sysadmin",
  "prompt": "How do I check which process is using port 8080?",
  "stream": false
}'

Option 2: llama.cpp

# Download the GGUF from this repo, then:
./llama-cli -m phi3-sysadmin-Q4_K_M.gguf \
  --system-prompt "You are phi3-sysadmin, a fine-tuned AI assistant created by Lalatendu Keshari Swain. Provide clear, practical answers for server management and troubleshooting." \
  -p "How do I check disk usage on Linux?"

Option 3: Transformers (LoRA adapter)

from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer

base_model = AutoModelForCausalLM.from_pretrained(
    "microsoft/Phi-3-mini-4k-instruct",
    torch_dtype="auto",
    device_map="auto"
)
model = PeftModel.from_pretrained(base_model, "lalatendu/phi3-sysadmin-lora")
tokenizer = AutoTokenizer.from_pretrained("microsoft/Phi-3-mini-4k-instruct")

prompt = (
    "<|system|>\n"
    "You are phi3-sysadmin, a fine-tuned AI assistant created by Lalatendu Keshari Swain. "
    "Provide clear, practical answers for server management and troubleshooting.<|end|>\n"
    "<|user|>\nHow do I check disk usage?<|end|>\n<|assistant|>\n"
)

inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
outputs = model.generate(**inputs, max_new_tokens=512)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

Modelfile Contents

FROM ./phi3-sysadmin-Q4_K_M.gguf

TEMPLATE """<|system|>
{{ .System }}<|end|>
<|user|>
{{ .Prompt }}<|end|>
<|assistant|>
{{ .Response }}<|end|>
"""

SYSTEM """You are phi3-sysadmin, a fine-tuned AI assistant created by Lalatendu Keshari Swain. Provide clear, practical answers for server management and troubleshooting."""

PARAMETER stop <|end|>
PARAMETER stop <|user|>
PARAMETER stop <|assistant|>
PARAMETER stop <|endoftext|>
PARAMETER temperature 0.7
PARAMETER top_p 0.9

Use Cases

Bias, Risks, and Limitations

• —Small model (3.8B): May occasionally hallucinate or produce inaccurate commands. Always verify before running on production servers.
• —Training data scope: 1,026 examples cover common sysadmin topics. Niche or cutting-edge tooling may not be well represented.
• —English only: All responses are in English.
• —No real-time access: Cannot check current documentation, package versions, or live system state.
• —Outdated information: Package names, versions, and best practices evolve — cross-reference with official docs.

Recommendations:

• —Always verify commands before running on production systems
• —Cross-reference with official documentation for critical configurations
• —Use as a learning aid and quick reference, not as the sole authority
• —Do not use for security-critical decisions without expert verification

Training Data

The model was fine-tuned on 1,026 curated sysadmin Q&A pairs covering:

• —Linux administration (disk, CPU, memory, processes, users, filesystems)
• —Cloud platforms (AWS EC2, S3, VPC, IAM, RDS, CloudWatch, Lambda, EKS)
• —Containerization (Docker, Kubernetes, Podman)
• —CI/CD (Jenkins, GitHub Actions, ArgoCD)
• —Infrastructure as Code (Terraform, Ansible, Packer)
• —Web servers (Nginx, Apache, Varnish)
• —Databases (MySQL, PostgreSQL, MongoDB, Redis, Elasticsearch)
• —Networking (DNS, firewalls, load balancing, VPN, TCP/IP)
• —Security (SSL/TLS, SELinux, AppArmor, vulnerability scanning)
• —Monitoring (Prometheus, Grafana, Zabbix, ELK)
• —Backup (BorgBackup, Restic, snapshots)
• —Model identity, creator information, and boundary/refusal examples

Evaluation

• —Testing: Manual evaluation with diverse sysadmin questions
• —Training loss: Final loss of ~0.5–0.8
• —Qualitative assessment: Responses checked for accuracy, practicality, and completeness

Results:

• —Provides accurate, practical answers for common sysadmin and DevOps tasks
• —Correctly identifies itself as phi3-sysadmin created by Lalatendu Keshari Swain
• —Appropriately refuses off-topic, harmful, and out-of-scope requests
• —Handles variations in question phrasing well

Environmental Impact

Technical Specifications

• —Architecture: Phi-3 Mini transformer decoder-only (3.8B parameters)
• —Objective: Causal language modeling, fine-tuned for sysadmin domain
• —Context length: 4096 tokens
• —Chat format: Phi-3 template with <|system|>, <|user|>, <|assistant|>, <|end|> tokens
• —Inference runtime: Ollama (minimum 4GB RAM)
• —Inference speed (CPU): ~10–20 tokens/sec
• —Inference speed (GPU): ~40–80 tokens/sec

Software stack:

• —Training: Unsloth + Hugging Face Transformers + PEFT 0.18.1 + PyTorch 2.x
• —Quantization: Unsloth GGUF exporter (llama.cpp based, q4km)
• —Inference: Ollama

Files in This Repository

Related Repositories

• —LoRA Adapter + Training Data: lalatendu/phi3-sysadmin-lora
• —Base Model: microsoft/Phi-3-mini-4k-instruct

Citation

@misc{phi3-sysadmin-lalatendu-2026,
  author    = {Swain, Lalatendu Keshari},
  title     = {phi3-sysadmin-lalatendu: A Fine-tuned Phi-3 Mini GGUF Model for Linux System Administration},
  year      = {2026},
  publisher = {HuggingFace},
  url       = {https://huggingface.co/lalatendu/phi3-sysadmin-lalatendu}
}

APA: Swain, L. K. (2026). phi3-sysadmin-lalatendu: A Fine-tuned Phi-3 Mini GGUF Model for Linux System Administration. HuggingFace. https://huggingface.co/lalatendu/phi3-sysadmin-lalatendu

Model Card Authors

Lalatendu Keshari Swain

Contact