model:
  name: FastText+BiLSTM
  architecture: BiLSTM
  embedding: pretrained FastText (subword)
  embedding_dim: 300
  hidden_dim: 256
  dropout: 0.5
  max_seq_len: 128
training:
  optimizer: Adam
  learning_rate: 0.001
  weight_decay: 0
  epochs: 10
  batch_size: 32
  loss: binary_crossentropy
data:
  dataset: PRISM (combined_hate_speech_dataset, Kaggle)
  target: hate_label
  classes: 2
  split: stratified 70/30 holdout, train pool 60/10 train/val
best_result:
  regime: langwise_regular
  strategy: english
  accuracy: 0.752355
  f1: 0.768844
  auc_roc: 0.825146