Gemma 4 E4B Mixer Assistant

This is Gemma 4 E4B (instruct), fine-tuned with LoRA. It turns a producer's or sound engineer's request into tool calls for an Ableton Live-style DAW. For example:

"Bring the kick down a touch, put a reverb on the snare and loop bars 5 to 8." β†’ set_volume(Kick, -1, relative), add_effect(Snare, Reverb), set_loop(start_bar 5, length 4)

It was made for the University of Michigan course Human-AI Interaction (Fall 2026, Assignment 2). The app, the dataset generator, the evaluation and the training code are on GitHub: ishaanjagyasi/Gemma4-DAW-VUI.

Files

File What it is
gemma4-mixer-Q4_K_M.gguf Merged model, 4-bit Q4_K_M (5.3 GB), for Ollama or llama.cpp
Modelfile Ollama Modelfile: Gemma 4 renderer and parser, so tool calls come back as structured tool_calls
lora/ The LoRA adapter (175 MB) for unsloth/gemma-4-E4B-it, for use with PEFT or Unsloth

Use with Ollama

hf download ishaanjagyasi/gemma4-e4b-mixer-assistant gemma4-mixer-Q4_K_M.gguf Modelfile --local-dir .
ollama create gemma4-mixer -f Modelfile

Then call /api/chat with the app's system prompt, the session as text, the 24 tool schemas, think: false and a context of at least 8k tokens. The prompt alone is about 3.5k tokens.

Training

  • Data: ishaanjagyasi/gemma4-mixer-tool-calls, 3,000 synthetic examples. Each is a session described as text plus a request, with the correct tool calls as the answer. They mix casual, spoken requests ("a touch", "the drums") with precise ones ("-7.5 dB", "126.5 BPM"), with 1–6 requests each.
  • Method: LoRA rank 16, alpha 16, on the language model's attention and MLP layers. The loss covers only the assistant's tool calls. Gemma 4's chat template, thinking off.
  • Schedule: 1 epoch, 750 steps of batch 4, learning rate 2e-4.
  • Hardware: 1 Γ— A40 for 6.5 h. Final training loss 0.0067.

Evaluation

Success means the session after the model's tool calls exactly matches the target session.

Untrained Gemma 4 E4B This model
Held-out validation (100) β€” 98%
Study set, standard (200) 47.0% 97.0%
Study set, hard (84): described tracks, comparisons, unseen slang, same-name tracks, multi-step, long batches 19.0% 65.5%

Limitations. The model was trained on generated requests only. It is weak at:

  • values defined relative to another track ("3 dB quieter than the kick");
  • telling apart tracks that share a name;
  • wording it has not seen.

Check its changes before keeping them.

Downloads last month
37
GGUF
Model size
8B params
Architecture
gemma4
Hardware compatibility
Log In to add your hardware

4-bit

Inference Providers NEW
This model isn't deployed by any Inference Provider. πŸ™‹ Ask for provider support

Model tree for ishaanjagyasi/gemma4-e4b-mixer-assistant

Adapter
(85)
this model

Dataset used to train ishaanjagyasi/gemma4-e4b-mixer-assistant