config.yaml

# Version 2: Improved Transformer
# With label smoothing, better optimizer, and beam search

# Data Configuration
data:
  src_lang: "vi"
  tgt_lang: "en"
  train_src: "data/raw_opus100/train.vi.txt"
  train_tgt: "data/raw_opus100/train.en.txt"
  # Validation files (if not exist, will auto-split from training)
  val_src: "data/raw_opus100/val.vi.txt"
  val_tgt: "data/raw_opus100/val.en.txt"
  val_split: 0.1  # 10% of training data for validation if val files don't exist
  test_src: "data/raw_opus100/public_test.vi.txt"
  test_tgt: "data/raw_opus100/public_test.en.txt"
  max_seq_length: 128

# Vocabulary
vocab:
  src_vocab_size: 40000  # Increased for better Vietnamese coverage
  tgt_vocab_size: 40000
  min_freq: 2

# Model - Same architecture
model:
  d_model: 512
  n_heads: 8
  n_encoder_layers: 6
  n_decoder_layers: 6
  d_ff: 2048
  dropout: 0.1
  max_seq_length: 512

# Training - Improved
training:
  batch_size: 32
  epochs: 8  # Good balance of quality and time
  optimizer: "adamw"  # Changed to AdamW
  learning_rate: 0.0001
  weight_decay: 0.01  # Added weight decay
  scheduler: "warmup"
  warmup_steps: 4000
  label_smoothing: 0.1  # Added label smoothing
  gradient_accumulation_steps: 2  # Effective batch size = 64
  max_grad_norm: 1.0
  
  use_wandb: true
  save_every: 1000
  eval_every: 500
  log_every: 100
  early_stopping_patience: 5

# Inference
inference:
  beam_size: 5  # Beam search
  max_decode_length: 128
  length_penalty: 0.6

# Paths
paths:
  checkpoint_dir: "experiments/v2_vi2en/checkpoints"
  log_dir: "experiments/v2_vi2en/logs"
  vocab_dir: "data/vocab_v2_vi2en"

device: "cuda"
seed: 42

# Weights & Biases
wandb:
  project: "nlp-transformer-mt"
  entity: null

# Version info
version:
  name: "v2_vi2en"
  description: "Improved Vi→En training with label smoothing and AdamW"
  improvements:
    - "Label smoothing (0.1)"
    - "AdamW optimizer with weight decay"
    - "Beam search (size=5)"
    - "Gradient accumulation"
    - "Early stopping"