JARVIS MARK VI - Neural Router

This is a fine-tuned FunctionGemma-270M model optimized for the JARVIS MARK VI ecosystem. It functions as a high-speed, offline intent classifier and tool router.

Capabilities

The model is trained to route queries to the following tools:

  • control_light: Controls smart lights - turn on, off, or dim lights in a room
  • web_search: Searches the web for information using Google
  • set_alarm: Sets an alarm for a specific time
  • add_task: Adds a task to the to-do list
  • get_system_info: Retrieves current system status including active timers, alarms, weather, and news
  • set_timer: Sets a countdown timer for a specified duration
  • create_calendar_event: Creates a new calendar event or appointment
  • read_calendar: Reads and retrieves calendar events for a date or time range
  • vision: Analyze what is visible on the screen or via the camera. Use this for questions like 'what is on my screen?', 'describe this image', 'what is this website about?', or 'look at the camera'.
  • open_app: Launches a specified application on the Windows OS. Supported apps include Chrome, VS Code, Notepad, Terminal, Calculator, and Spotify.
  • close_app: Terminates a running application on the Windows OS.
  • set_vision_mode: Enables or disables specialized vision modes like hand tracking or gesture recognition.
  • passthrough: DEFAULT FUNCTION - Use this whenever no other function is clearly needed. This is the fallback for: greetings (hello, hi, good morning), chitchat (how are you, what's your name), general knowledge questions, explanations, conversations, and ANY query that does NOT explicitly require controlling lights, setting timers, searching the web, or managing calendar events. When in doubt, use passthrough.

Technical Details

  • Base Model: google/functiongemma-270m-it
  • Training Data: Synthetic Stark-Augmented Dataset
  • Format: call:function_name{parameters}
  • Optimization: LoRA (Rank 32) merged into base weights.

Usage

This model is intended for use with the JARVIS MARK VI routing engine.

Downloads last month
1
Safetensors
Model size
0.3B params
Tensor type
BF16
·
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support