data:
  language: ne
  sample_rate: 16000
  hf_dataset_id: Firoj112/voxcpm-nepali-data
  train_manifest: finetuning_data/manifests/train_manifest.jsonl
  val_manifest: finetuning_data/manifests/val_manifest.jsonl
hardware:
  gpu_count: 1
  load_in_4bit: false
  gradient_checkpointing: true
  vram_gb_per_gpu: 24.0
lora:
  alpha: 96
  bias: none
  dropout: 0.05
  enabled: true
  rank: 96
  target_modules:
  - q_proj
  - v_proj
  - k_proj
  - o_proj
metadata:
  config_type: lora
  language: ne
  model_version: '2'
model:
  name_or_path: openbmb/VoxCPM2
  optimize: false
  version: '2'
training:
  amp: bf16
  bf16: true
  dataloader_num_workers: 0
  ddp_find_unused_parameters: false
  epochs: 20
  eval_steps: 500
  fp16: false
  gradient_accumulation_steps: 32
  gradient_checkpointing: true
  learning_rate: 0.00005
  logging_steps: 50
  max_grad_norm: 1.0
  max_steps: 20000
  output_dir: ./checkpoints/lora_nepali_v2
  per_device_train_batch_size: 1
  report_to:
  - wandb
  - tensorboard
  save_steps: 500
  warmup_steps: 500
  lr_scheduler_type: cosine
wandb:
  log_model: checkpoint
  project: Vox-CPM-Nepali-Fine-Tune
  watch: gradients