9FA/baharani_v12_lora

Baharani v12 VoxCPM2 LoRA

LoRA fine-tune of openbmb/VoxCPM2 on wldaldakheel/baharani-mix.

v12 Training Recipe

Param	Value
LoRA rank	64
LoRA alpha	128 (ratio 2.0)
LoRA dropout	0.0 (implicit regularization from rank constraint)
Learning rate	5e-5
Scheduler	Cosine with 1000-step warmup
Steps	15000 (~3.1 epochs)
Batch size	4 × 4 grad_accum = 16 effective
EMA decay	0.999
Validation	Every 3000 steps, best checkpoint selected

Key improvements over v10

• LoRA rank 64/alpha=128 (was r=32/alpha=32): more capacity for dialect features
• Dropout removed: 68h of data provides natural regularization; dropout hurts audio quality
• 3.1 epochs (was ~1 epoch): proper convergence without underfitting
• Cosine LR: smooth decay instead of constant → better convergence
• EMA decay=0.999: appropriate halflife for 15K training steps
• Train/val split: best checkpoint selection by validation loss
• Lower LR (5e-5 vs 1e-4): finer adaptation without catastrophic forgetting

Usage

For EMA weights (recommended for production):

License

Apache 2.0