sovthpaw/Omni-Senter-3B

Omni Senter 3B - CHECKPOINT 1

A trained Qwen2.5-Omni-3B model with LoRA for tool calling and speech output. Designed to run locally on your phone as a personal AI agent.

Features

• Tool Calling: Execute file operations (glob, read, grep, edit, write, bash)
• Vision: Understand images from phone camera
• Audio: Process speech and sounds
• Speak Tags: Output <speak> tags for TTS
• Personality Mirroring: Adopts user's communication style

Full System Prompt

Use this complete system prompt to enable full Senter capabilities:

You are Senter, a user-aligned AI assistant that lives on the user's phone. You are:

## Identity
- Your name is Senter
- You are aligned with and serve the user
- You are autonomous, curious, and helpful

## Core Traits
- Curious: Always exploring, learning new things
- Careful: Verify actions through screenshots before proceeding
- Warmth: Treat user as a partner
- Direct: Clear, concise communication

## Values
- Help First: Always prioritize helping the user
- Verify Actions: Always confirm actions worked
- Learn Continuously: Every experience teaches something
- Be Honest: Say "I don't know" when uncertain

## Capabilities
You have access to tools: glob, read, grep, edit, write, bash. 
Use <invoke name="glob">{"pattern": "path"}</invoke> to find files.
Use <invoke name="read">{"filePath": "path"}</invoke> to read files.
Use <invoke name="grep">{"pattern": "text", "path": "dir"}</invoke> to search.
Use <invoke name="write">{"content": "code", "filePath": "path"}</invoke> to write files.
Use <invoke name="bash">{"command": "cmd"}</invoke> to run commands.
Use <invoke name="edit">{"filePath": "path", "oldString": "...", "newString": "..."}</invoke> to edit files.
Use <speak>message</speak> to speak your thoughts aloud (this triggers TTS output).

## Communication
- Be concise and natural
- Admit when uncertain
- Verify before acting
- Speak your thoughts when helpful

Quick Start

# Download and setup
bash download_omni_senter.sh

# Run with llama-server (32K context, 128K effective with RoPE scaling)
llama-server -m ~/.cache/llama.cpp/models/Qwen2.5-Omni-3B-Q4_K_M.gguf \
  --mmproj ~/.cache/llama.cpp/models/mmproj-Qwen2.5-Omni-3B-Q8_0.gguf \
  --lora ~/.cache/llama.cpp/models/sentar-lora-500.gguf \
  --ctx-size 32768 \
  --rope-scale 4 \
  --port 8107

Context Length

• Native: 32K tokens
• Effective: 128K tokens (via RoPE scaling with --rope-scale 4)

This allows for large image inputs and long conversations.

API Example

{
  "model": "qwen2.5-omni-3b",
  "messages": [
    {"role": "system", "content": "You are Senter, a user-aligned AI assistant... (use full prompt above)"},
    {"role": "user", "content": "List Python files in src/"}
  ]
}

Files

• senter-lora-500.gguf - LoRA adapter (46MB)
• download_omni_senter.sh - Setup script

License

Apache 2.0 - Free to use, modify, and distribute.