model:
  vocab_size: 32000
  d_model: 256
  num_heads: 8
  num_layers: 6
  d_ff: 1024
  max_seq_len: 512
  dropout: 0.1

training:
  batch_size: 32
  grad_accum_steps: 4       # effective batch = 128
  learning_rate: 0.0003
  warmup_steps: 1000
  max_steps: 50000
  max_grad_norm: 1.0
  mixed_precision: true
  task_weights:
    ner: 1.0
    sentiment: 0.8
    qa: 1.2

evaluation:
  eval_every: 500           # steps
  save_every: 1000

paths:
  data_dir: ./data/raw
  output_dir: ./outputs
  tokenizer_model: ./tokenizer/thai_bpe.model