burrak/moshi-burak-fine-tune

Moshi Fine-tuned LoRA - Burak's Voice AI

This is a LoRA adapter fine-tuned on Moshi to answer questions about Burak's background and experience.

Model Details

• Base Model: kyutai/moshiko-pytorch-bf16
• Training Method: LoRA (Low-Rank Adaptation)
• LoRA Rank: 16
• Training Steps: 500
• Dataset: 223 Q&A conversations (~47 minutes)

Usage

# Install Moshi
pip install git+https://github.com/kyutai-labs/moshi.git#subdirectory=moshi

# Run with LoRA
python -m moshi.server \
  --lora-weight=lora.safetensors \
  --config-path=config.json

Then open http://localhost:8998 in your browser.

Training Details

• Framework: moshi-finetune
• GPUs: 2x Kaggle T4
• Training Time: ~15 hours
• Batch Size: 1 per GPU (effective: 2)

License

Apache 2.0 (same as base Moshi model)