project:
  name: "hate-speech-detector"
  version: "0.1.0"
  seed: 42

data:
  # Dataset paths
  davidson_path: "data/raw/davidson_english.csv"
  russian_path: "data/raw/russian_toxic.csv"

  # Preprocessing
  max_length: 128
  num_classes: 2
  class_names: ["non-toxic", "toxic"]

  # Splits
  train_size: 0.7
  val_size: 0.15
  test_size: 0.15

model:
  name: "xlm-roberta-base"
  num_labels: 2
  dropout: 0.1

training:
  # ЛУЧШИЕ ПАРАМЕТРЫ ИЗ GRID SEARCH:
  batch_size: 16
  learning_rate: 1e-5
  num_epochs: 5  # полное обучение
  warmup_steps: 500
  weight_decay: 0.01

  # Для дисбаланса
  use_class_weights: true

  # Early stopping
  patience: 3

paths:
  data_raw: "data/raw"
  data_processed: "data/processed"
  models: "models"
  logs: "logs"
  mlruns: "mlruns"

mlflow:
  experiment_name: "bilingual-hate-speech"
  tracking_uri: "http://localhost:5000"