zk-mohammad/cluster-worker-v2-0.5b-GGUF

Cluster Worker v2 0.5B GGUF

This is a trailer — contact for the fully upgraded version with enhanced fine-tuning and extended coverage.

Need a custom model for your specific task? contact. mohammadzaidkhanofficial@gmail.com

Cluster Worker v2 is a highly specialized Small Language Model (SLM) fine-tuned on 65,000+ Linux & DevOps task samples. It is engineered specifically to act as a deterministic, zero-fluff CLI execution worker inside dual-agent AI architectures.

Unlike standard chat models, this model uses response-only loss masking to completely eliminate conversational filler, explanations, and unnecessary Markdown wrappers — returning strictly valid, executable Bash commands.

---

Licensing & Commercial Use

This model is released under the Business Source License 1.1 (BSL 1.1).

• Free Use: Free for non-production, testing, academic, and evaluation purposes.
• Commercial Use: If you wish to use this model in production for commercial purposes, applications, or enterprise services, you must acquire a commercial license.

Contact for Enterprise Licensing

For production keys, custom fine-tuning on your company data, or commercial licensing inquiries, please reach out directly:

• Email: mohammadzaidkhanofficial@gmail.com
• LinkedIn: https://www.linkedin.com/in/mohammad-zaid-5021ab3b1

---

Key Features

• Zero Conversational Fluff: No "Sure, here is your command:" or introductory text. It outputs raw code meant directly for stdout/terminal execution.
• Enterprise Sysadmin Coverage: Fine-tuned on multi-step Linux administrative tasks, file system manipulation, log parsing (awk, grep, sed), networking, and systemd management.
• Ultra-Low Footprint: Built on Qwen2.5-Coder-0.5B-Instruct and quantized to Q8_0, taking up under 600 MB of VRAM/RAM.
• Optimized for Dual-Agent Architectures: Serves as a zero-cloud-token execution node when paired with primary reasoning models (e.g., Gemini, Claude, GPT-4o, or a 7B Architect model).

---

---

Quickstart & Deployment

Method 1: Automated Script (Recommended)

If using Debian/Ubuntu Linux, run the official automated setup script to install Ollama, build the local wrapper, and launch the worker:

curl -fsSL https://raw.githubusercontent.com/cl-andro/cluster-ai-training-sets/main/cluster-worker-nonthink-ai.sh | bash

---

Method 2: Ollama (Modelfile)

You can run this model directly in Ollama using the ChatML standard format.

1. Create a file named Modelfile:

FROM hf.co/zk-mohammad/cluster-worker-v2-0.5b-GGUF

TEMPLATE """<|im_start|>system
You are a strict, zero-fluff Linux terminal worker. Return strictly the bash command, no explanations, no markdown chat, no formatting.<|im_end|>
<|im_start|>user
{{ .Prompt }}<|im_end|>
<|im_start|>assistant
"""

PARAMETER temperature 0.0
PARAMETER top_k 1
PARAMETER stop "<|im_end|>"
PARAMETER stop "<|im_start|>"

2. Build and run the model:

ollama create worker -f Modelfile
ollama run worker "Find all files ending in .log in /var/log older than 7 days and delete them."

---

Method 3: llama.cpp CLI

Run directly via llama-cli with Jinja template support enabled:

llama-cli -hf zk-mohammad/cluster-worker-v2-0.5b-GGUF \
  --jinja \
  --temp 0.0 \
  -p "Find all files ending in .log in /var/log older than 7 days and delete them."

---

Example Outputs

User Request	Worker Model Output
Find all .log files in /var/log older than 7 days and delete them.	find /var/log -name '*.log' -mtime +7 -delete
Extract column 1 from access.log, count unique values, and print top 5.	awk '{print $1}' access.log | sort | uniq -c | sort -rn | head -5
List all hidden files with human-readable file sizes.	ls -la

---

Prompt Format (ChatML)

This model follows the standard Qwen2.5 ChatML prompt structure:

<|im_start|>system
You are a strict, zero-fluff Linux terminal worker. Return strictly the bash command, no explanations, no markdown chat, no formatting.<|im_end|>
<|im_start|>user
[Your Terminal Task Request]<|im_end|>
<|im_start|>assistant

---

Recommended Inference Parameters

To ensure deterministic execution and prevent conversational drift:

• Temperature: 0.0
• Top_K: 1
• Sampling: Disabled (Greedy decoding)

---