model:
  name: Word2Vec+BiLSTM
  architecture: BiLSTM
  embedding: pretrained Word2Vec
  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: multistage
  strategy: combined
  accuracy: 0.730456
  f1: 0.698889
  auc_roc: 0.809126