Denniscor/musicgen-mara

MusicGen Mara Beat (LoRA)

Fine-tuned MusicGen-small on Nigerian mara/afrobeat dance instrumentals.

Training

• Method: LoRA (r=16, alpha=32) on decoder attention
• Dataset: Denniscor/mara-beats -- 73 vocal-free clips, 35 min
• Epochs: 30
• Best loss: 7.1116

Usage

from transformers import MusicgenForConditionalGeneration, AutoProcessor
from peft import PeftModel
import scipy

base = MusicgenForConditionalGeneration.from_pretrained("facebook/musicgen-small")
model = PeftModel.from_pretrained(base, "Denniscor/musicgen-mara")
processor = AutoProcessor.from_pretrained("facebook/musicgen-small")

inputs = processor(
    text=["nigerian mara dance beat with heavy percussion and afrobeat drums"],
    return_tensors="pt",
)
audio = model.generate(**inputs, max_new_tokens=1500)  # ~30 seconds
sr = model.config.audio_encoder.sampling_rate
scipy.io.wavfile.write("mara_beat.wav", sr, audio[0, 0].cpu().numpy())