metadata

title: LLM Forest Orchestra
emoji: 📈
colorFrom: green
colorTo: green
sdk: gradio
sdk_version: 5.44.1
app_file: app.py
pinned: true

title: Locutusque Models emoji: 🌲 colorFrom: green colorTo: blue sdk: gradio sdk_version: 4.29.0 app_file: app.py pinned: false license: apache-2.0

LLM Forest Orchestra — Hugging Face Space (Gradio)

This Space turns transformer hidden states and attentions into a layered MIDI composition.

Features

Two compute modes:
- Full model: loads a base HF model (default: unsloth/Qwen3-14B-Base) and extracts internals.
- Mock latents: fast CPU-friendly demo mode that synthesizes tensors to preview the music system.
Musical controls: scale selection (with custom notes), ticks-per-beat grid, velocity range, and instrument/role presets.
Exports a .mid file you can drop into any DAW.

Files

app.py — Gradio app entry.
requirements.txt — Python deps.
README.md — You are here.

Hardware / tips

unsloth/Qwen3-14B-Base is large. On CPU-only Spaces, use Mock latents. For Full model, pick a smaller base or provision a GPU Space.
Models must support output_hidden_states=True and output_attentions=True.
MIDI channels are limited to 16; the UI caps layers at 6 for headroom.

Inspiration

This project is inspired by the way mushrooms and mycelial networks in forests connect plants and trees, forming a living web of communication and resource sharing. These connections, can be turned into ethereal music. Just as signals move through these hidden connections, transformer models also pass hidden states and attentions across their layers. Here, those hidden connections are translated into music, analogous to the forest's secret orchestra.