# Transformer Model Configuration
model:
  name: myanmar_ghost_transformer
  type: transformer
  architecture: custom

# Model architecture
architecture:
  hidden_size: 768
  num_layers: 12
  num_heads: 12
  intermediate_size: 3072
  dropout: 0.1
  attention_dropout: 0.1

# Vocab
vocab:
  type: sentencepiece
  vocab_size: 32000
  special_tokens:
    pad: "<pad>"
    unk: "<unk>"
    bos: "<s>"
    eos: "</s>"

# Training
training:
  batch_size: 16
  learning_rate: 5e-5
  weight_decay: 0.01
  adam_beta1: 0.9
  adam_beta2: 0.999
  gradient_accumulation_steps: 4
  max_grad_norm: 1.0
  num_epochs: 10
  warmup_steps: 500
  scheduler: linear

# Mixed precision
mixed_precision: true
fp16: true

# Regularization
regularization:
  dropout: 0.1
  label_smoothing: 0.1
  early_stopping:
    enabled: true
    patience: 3
    monitor: val_loss