🤖 NanoHat-360M

A lightweight Fedora Linux desktop AI agent fine-tuned from SmolLM2-360M-Instruct for multi-turn system interaction, structured tool calling, and Linux desktop assistance.

Project: NanoHat

The model is designed to operate through the NanoHat runtime, which provides the smolagents execution harness, FedoraFormatBridge, six consolidated tools, and security-hardened system execution.

✨ Highlights

• Fedora/Linux-focused desktop assistance
• Structured <thought> and <tool_call> generation
• Multi-tool diagnostic and system workflows
• Advanced Fedora/Linux sysadmin coverage
• Semantic cross-run dataset deduplication
• Canonical JSON validation and auto-rescue
• Argument alias and synonym normalization
• Assistant-only loss masking during fine-tuning
• Security-focused subprocess execution

🛠️ Supported Tools

Tool	Purpose
calculator(expression)	Mathematical evaluations
web_search(query)	Live web search queries
user_memory(action, key, value)	Persistent user preference storage
reminder(task, time_or_delay)	Desktop notifications and delayed reminders
system_health(target)	CPU, RAM, disk, process, battery, and system telemetry
system_action(action, target)	Controlled OS actions including Wi-Fi, Bluetooth, application launching, screenshots, and screen locking

🐧 Fedora/Linux Coverage

The training curriculum covers a broad range of desktop and system-administration workflows, including:

• SELinux and ausearch
• firewalld configuration
• SSH keys, forwarding, and host aliases
• Git rebase/stash conflict recovery
• Python PEP 668 / virtual environments
• Btrfs snapshots and LVM storage
• GRUB, suspend, and power troubleshooting
• Wayland, PipeWire, OBS Studio, and fractional scaling
• Wacom tablets, USB-C docks, and webcams
• DNS, SMB/NFS, and NetworkManager configuration
• dnf5, system upgrades, COPR, and rpm-ostree
• Bluetooth, audio, VPN, Podman, multi-monitor, and other desktop troubleshooting scenarios

🧠 Dataset & Training

The model was fine-tuned on a curated 1,094-sample golden dataset containing validated multi-turn Fedora/Linux interaction trajectories.

The dataset was generated through a structured curriculum covering:

• Single-tool interactions
• No-tool conversations
• Multi-tool workflows
• Edge cases and recovery scenarios
• Advanced Fedora/Linux system-administration workflows

Multiple generation runs were consolidated into a unified golden dataset using semantic cross-run deduplication and canonical validation.

The validation pipeline includes Fedora-grounded tool outputs, novelty prompting, JSON syntax auto-rescue, argument synonym normalization, turn-structure correction, and deterministic JSON re-serialization.

🔐 Runtime Safety

NanoHat uses a security-hardened subprocess execution model:

• No os.system()
• No shell=True
• Discrete subprocess argument lists
• Process-name validation
• Confirmation gates for destructive operations
• Explicit Bluetooth state checks before power changes

The model itself should not be treated as a security boundary. Runtime-side validation and authorization remain responsible for enforcing safe system actions.

🚀 Quickstart

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "blackpirates/NanoHat-360M"

tokenizer = AutoTokenizer.from_pretrained(model_id)

model = AutoModelForCausalLM.from_pretrained(
    model_id,
    torch_dtype=(
        torch.bfloat16
        if torch.cuda.is_bf16_supported()
        else torch.float32
    ),
    device_map="auto",
)

messages = [
    {
        "role": "system",
        "content": (
            "You are a helpful AI agent running on Fedora Linux. "
            "Available tools: [calculator, web_search, user_memory, "
            "reminder, system_health, system_action]."
        ),
    },
    {
        "role": "user",
        "content": "Check my CPU usage and set a reminder to drink water in 20 minutes.",
    },
]

input_text = tokenizer.apply_chat_template(
    messages,
    tokenize=False,
    add_generation_prompt=True,
)

inputs = tokenizer(input_text, return_tensors="pt").to(model.device)

outputs = model.generate(
    **inputs,
    max_new_tokens=256,
    temperature=0.2,
)

print(
    tokenizer.decode(
        outputs[0][inputs.input_ids.shape[1]:],
        skip_special_tokens=True,
    )
)

📊 Training Details

• Base Model: HuggingFaceTB/SmolLM2-360M-Instruct
• Method: QLoRA
• LoRA Rank: 16
• LoRA Alpha: 32
• Target Modules: All linear projections
• Loss Masking: Loss computed exclusively on assistant-generated turns (<thought>, <tool_call>, and assistant responses)
• Epochs: 3
• Final Evaluation Loss: 0.8567

⚠️ Limitations

This is a 360M-parameter specialized agent model, not a general-purpose frontier language model.

Performance may degrade on:

• Unseen operating systems
• Complex general reasoning tasks
• Tasks outside the trained toolset
• Unsupported tool arguments or workflows
• Long-horizon autonomous tasks

For real system operations, always validate model-generated actions at the runtime/tool layer before execution.