BAP-Labs-M1

LoRA adapter for Hermes-2-Pro-Mistral-7B fine-tuned to generate Serum synthesizer preset parameters from natural language descriptions.

Model Details

• Base Model: NousResearch/Hermes-2-Pro-Mistral-7B
• Fine-tuning Method: LoRA (Low-Rank Adaptation)
• Training Framework: MLX-LM (Apple Silicon optimized)
• Task: Text-to-Synthesizer-Parameters generation

Training Configuration

• Dataset: 897 examples of natural language descriptions paired with Serum preset parameters
• Training Split: 90/10 (807 train / 90 validation)
• Iterations: 300
• Batch Size: 8
• Learning Rate: 3e-4
• LoRA Layers: 16
• Trainable Parameters: 1.704M (0.024% of base model)

Training Results

• Final Train Loss: 0.782
• Final Validation Loss: 0.757
• Training Time: ~5.5 hours on M4 Max
• Peak Memory: 62.8 GB

Usage

With MLX-LM

from mlx_lm import load, generate

# Load model with LoRA adapter
model, tokenizer = load(
    "NousResearch/Hermes-2-Pro-Mistral-7B",
    adapter_path="bapinero/BAP-Labs-M1"
)

# Generate Serum preset parameters
prompt = """<|im_start|>system
You are a Serum synthesizer preset designer. Generate JSON parameter changes for Serum presets based on natural language descriptions.<|im_end|>
<|im_start|>user
Create a deep dubstep bass with lots of wobble<|im_end|>
<|im_start|>assistant
"""

response = generate(model, tokenizer, prompt=prompt, max_tokens=500, temp=0.7)
print(response)

Command Line

mlx_lm.generate \
  --model NousResearch/Hermes-2-Pro-Mistral-7B \
  --adapter-path bapinero/BAP-Labs-M1 \
  --prompt "Create a bright future bass lead" \
  --max-tokens 500 \
  --temp 0.7

Output Format

The model generates JSON with Serum parameter changes:

{
  "parameter_changes": [
    {"parameter_index": 22, "new_value": 0.85},
    {"parameter_index": 77, "new_value": 0.12},
    ...
  ]
}

Limitations

• Trained specifically for Serum synthesizer (448 mapped parameters)
• Best results with genre-specific descriptions (dubstep, house, bass music, etc.)
• Optimized for MLX framework (Apple Silicon)

License

Based on Hermes-2-Pro-Mistral-7B. Please refer to base model license.

Citation

@misc{bap-labs-m1,
  author = {BAP Labs},
  title = {BAP-Labs-M1: Serum Synthesizer Control via LLM},
  year = {2025},
  publisher = {HuggingFace},
  url = {https://huggingface.co/bapinero/BAP-Labs-M1}
}