# Clean-Signal Dual-Input Hybrid — raw Text (BERT) + clean_text/metadata (LR)

pipeline:
  random_state: 42
  test_size: 0.2
  val_size: 0.15
  max_train_test_gap: 0.05
  target_f1_weighted: 0.80

data:
  raw_path: data/raw/youtoxic_english_1000.csv
  processed_preprocessed: data/processed/v2/comments_preprocessed.csv
  processed_stats: data/processed/v2/comments_with_stats.csv
  target_binary: IsToxic
  text_column: Text
  id_column: CommentId
  features_config: configs/features.yaml

augmentation:
  enabled: true
  strategy: back_translation
  source_lang: en
  pivot_lang: de
  min_words: 3
  max_words: 60
  rate_limit_every: 50
  rate_limit_sleep_sec: 1.0
  dedup:
    enabled: true
    cosine_threshold: 0.95
    embedding_model: sentence-transformers/all-MiniLM-L6-v2

transformer:
  model_id: unitary/toxic-bert
  max_length: 128
  reuse_checkpoint: models/expert_toxic_bert
  fixed_threshold: 0.33
  train_if_missing: false

logistic_regression:
  C: 0.05
  max_iter: 2000
  class_weight: balanced
  solver: lbfgs
  gap_search:
    enabled: true
    max_gap: 0.05
    use_original_train_for_gap: true
    param_grid:
      - {C: 0.05, max_features: 800, min_df: 3}
      - {C: 0.03, max_features: 500, min_df: 5}
      - {C: 0.02, max_features: 400, min_df: 5}
      - {C: 0.01, max_features: 400, min_df: 8}
      - {C: 0.005, max_features: 300, min_df: 10}
      - {C: 0.002, max_features: 250, min_df: 12}
  tfidf:
    max_features: 800
    ngram_range: [1, 2]
    sublinear_tf: true
    min_df: 3

ensemble:
  weight_metric: f1_weighted
  min_lr_weight: 0.15
  max_lr_weight: 0.45
  threshold_tuning:
    enabled: true
    metric: f1_weighted

output:
  lr_path: models/hybrid_clean_lr.joblib
  ensemble_meta_path: models/hybrid_clean_ensemble_meta.json
  reports_dir: reports/hybrid_clean