sonia-model-catalog

Sonia Model Catalog

This repository provides a lightweight, structured catalog of recommended GGUF models for use with the Sonia on-device music generation engine.
It does not contain any model weights — only metadata and download links to publicly available models hosted on HuggingFace.

Purpose

The Sonia app allows users to download and run local Large Language Models (LLMs) directly on their device for offline symbolic music generation.
This catalog serves as a centralized, versioned list of supported models, including:

• Model names
• Official source URLs
• Recommended quantizations
• File sizes
• SHA‑256 checksums (for integrity verification)

The app reads this catalog to present available models to the user and manage downloads.

File Structure

catalog.json   # List of supported GGUF models with URLs and metadata
README.md      # You are here

Example Entry (from catalog.json)

{
  "id": "Phi-3.5-mini-instruct",
  "name": "Phi-3.5 Mini Instruct (1.6B)",
  "description": "Fast and lightweight model suitable for on-device symbolic music generation.",
  "size_mb": 700,
  "url": "https://huggingface.co/microsoft/Phi-3.5-mini-instruct-gguf/resolve/main/Phi-3.5-mini-instruct-Q4_K_M.gguf",
  "sha256": "REPLACE_WITH_SHA256"
}

How Sonia Uses This Catalog

1. Fetches the file catalog.json from this repository.
2. Displays the available models to the user.
3. Downloads the selected GGUF file from its official source URL.
4. Validates file integrity using the provided SHA‑256 hash.
5. Loads the model via the on-device Llama.cpp backend.

This approach ensures:

• The app remains lightweight
• Models are always up to date
• Users choose the model best suited to their device
• No redistribution of model weights within this repo

Updating the Catalog

To update or add models:

1. Edit catalog.json
2. Add or update model entries
3. Provide accurate SHA‑256 checksums
4. Commit and push

Sonia automatically picks up new models without requiring an app update.

License

This repository contains metadata only and no model weights.
All referenced models remain under their respective licenses hosted on HuggingFace.