model:
  name: Many-to-one LSTM
  architecture: LSTM (many-to-one)
  embeddings_tested: [Word2Vec, GloVe, FastText]
  embedding_dim: 100
  hidden_dim: 128
  dropout: 0.2
  max_seq_len: 100
training:
  optimizer: Adam
  learning_rate: 0.01
  weight_decay: 0
  epochs: 30
  batch_size: 16
  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
results_regular_combined:
  Word2Vec+LSTM: {accuracy: 0.6675, f1: 0.5892, auc_roc: 0.7294}
  GloVe+LSTM: {accuracy: 0.6779, f1: 0.5644, auc_roc: 0.7532}
  FastText+LSTM: {accuracy: 0.6610, f1: 0.5729, auc_roc: 0.7208}