seed: 4

dataset:
  train_csv: data/raw/train.csv
  dev_csv: data/raw/dev.csv
  test_csv: data/raw/test.csv

preprocess:
  lowercase: false

classical:
  max_features: 20000

bilstm:
  max_len: 100
  min_freq: 2
  emb_dim: 128
  hidden_dim: 128
  batch_size: 256
  epochs: 50
  lr: 0.001
  dropout: 0.5
  optimizer: adam

phobert:
  model_name: vinai/phobert-base-v2
  text_source: raw_text
  max_len: 100
  batch_size: 16
  eval_batch_size: 32
  epochs: 4
  lr: 2.0e-5
  weight_decay: 0.01

phobert_bilstm:
  text_source: raw_text
  freeze_encoder: false
  head_type: cls_mlp
  max_len: 100
  hidden_dim: 256
  dropout: 0.5
  batch_size: 16
  epochs: 4
  encoder_lr: 1.0e-5
  head_lr: 3.0e-5
  llrd: 0.9
  warmup_ratio: 0.1
  weight_decay: 0.01
  grad_clip: 1.0
  loss_type: class_weight
  early_stopping_patience: 2
  threshold_grid: [0.4, 0.45, 0.5, 0.55, 0.6]

outputs:
  model_table_csv: results/tables/table_4_5_proposed_main.csv
  metrics_json: results/metrics/model_comparison_table_4_5.json
  log_txt: results/logs/table_4_5_model_comparison.log
  registry_csv: experiments/registry.csv
  proposed_checkpoint_pt: results/checkpoints/phobihsd_proposed.pt