hamma-gemma-4-devops-GGUF

HAMMA DevOps AI (Gemma 4 Q4_K_M) ⬛⬜

This is a specialized, locally-running DevOps and Sysadmin AI, fine-tuned specifically to diagnose and resolve Linux server errors, reverse proxy (Nginx) crashes, and common deployment bottlenecks.

This model serves as the core intelligence engine for HAMMA, an offline, Zero-Install SSH and SFTP client featuring a geometric brutalist interface. It was created as an official submission for the Gemma 4 Good Hackathon.

🧠 Model Details

• Base Model: unsloth/gemma-4-e4b-it-bnb-4bit
• Architecture: Gemma 4 (Text-only, 8B parameters)
• Quantization: Q4_K_M (4-bit, highly optimized for standard laptop hardware)
• File Size: 5.34 GB (Requires ~5.5GB to 6GB RAM to run natively)
• License: Gemma License

🎯 Intended Use & Out-of-Scope

Intended Use: - Assisting developers with Linux server troubleshooting (Ubuntu/Debian focused).

• Generating rapid, accurate terminal commands for system administration (e.g., port management, SSH configuration, Docker container debugging).
• Running entirely offline in low-connectivity environments to preserve maximum data privacy.

Out-of-Scope:

• General conversational chat or creative writing.
• Writing full application source code from scratch.
• Acting as a substitute for human oversight in production environments (always verify commands before execution).

📊 Training Data & Methodology

The model was fine-tuned to prioritize direct, action-oriented terminal commands over conversational padding.

• Dataset: Fine-tuned on a custom, curated dataset of 1,500+ problem-solution pairs. The data was synthesized from standard Ubuntu sysadmin man pages, Nginx documentation, and common systemd failure states, specifically formatted to enforce zero-fluff responses.
• Fine-Tuning Framework: LoRA (Low-Rank Adaptation) using Unsloth.
• Hyperparameters: - LoRA Rank ($r$): 16
  • LoRA Alpha ($\alpha$): 32
  • Target Modules: q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj
  • Epochs: 1 (Note: A single epoch was strictly sufficient to enforce the desired CLI-output format given the highly narrow task domain, preventing overfitting on the specific synthetic log errors).

⚠️ Limitations & Bias

• Destructive Commands: While the model is trained to fix errors, users must exercise caution. The model may occasionally suggest aggressive fixes (e.g., kill -9 or recursive rm commands) if prompted carelessly.
• Context Blindness: The AI does not know your specific server's custom configuration or architecture unless explicitly provided in the prompt.
• Hallucination: Like all LLMs, it can occasionally hallucinate flags for obscure CLI tools.

🏆 Evaluation

During informal benchmarking against 50 common Linux sysadmin failure states (including port collisions, dead systemd services, and SSH key rejection), this fine-tuned model consistently delivered the exact resolution command.

Methodology Note: The fine-tuned HAMMA model achieved a 3x faster time-to-generate the full response compared to the base gemma-4-e4b-it model. Because the base model tends to generate paragraphs of conversational padding ("I'd be happy to help you with that error..."), our fine-tuned adapter immediately outputs the raw terminal command, drastically cutting down total inference time.

💻 Usage (llama.cpp)

To run this model directly via the command line using llama.cpp. Note the flags: -c 4096 allocates a healthy context window for pasting large error logs, while -n 512 caps the generated response to keep the output concise.

llama-cli -m gemma-4-e4b-it.Q4_K_M.gguf -c 4096 -n 512 --prompt "Server error: 502 Bad Gateway on Nginx. Fix it."