project_name: security_nlp_3task
tasks:
- name: fraud_detection
  type: single_label_classification
  data_path: fixtures/sample_classification.jsonl
  heads:
  - name: fraud
    num_labels: 2
    weight: 1.0
    dropout: 0.1
  text_column: text
  label_column: label
  label_maps: null
- name: sentiment_analysis
  type: single_label_classification
  data_path: fixtures/sample_sentiment.jsonl
  heads:
  - name: sentiment
    num_labels: 3
    weight: 1.0
    dropout: 0.1
  text_column: text
  label_column: label
  label_maps: null
- name: call_quality
  type: question_answering
  data_path: /content/JengaAI/fixtures/synthetic_qa_metrics_data_v01x.json
  heads:
  - name: opening
    num_labels: 1
    weight: 1.0
    dropout: 0.1
  - name: listening
    num_labels: 5
    weight: 1.5
    dropout: 0.1
  - name: proactiveness
    num_labels: 3
    weight: 1.0
    dropout: 0.1
  - name: resolution
    num_labels: 5
    weight: 2.0
    dropout: 0.1
  - name: hold
    num_labels: 2
    weight: 0.5
    dropout: 0.1
  - name: closing
    num_labels: 1
    weight: 1.0
    dropout: 0.1
  text_column: transcript
  label_column: labels
  label_maps:
    opening:
      0: greeting
    listening:
      0: acknowledgment
      1: empathy
      2: clarification
      3: active_listening
      4: patience
    proactiveness:
      0: initiative
      1: follow_up
      2: suggestions
    resolution:
      0: identified_issue
      1: provided_solution
      2: confirmed_resolution
      3: set_expectations
      4: offered_alternatives
    hold:
      0: asked_permission
      1: explained_reason
    closing:
      0: proper_farewell
model:
  base_model: roberta-base
  hidden_size: 768
  dropout: 0.1
  fusion:
    type: attention
    dropout: 0.1
    use_residual: true
    num_attention_heads: 1
    gate_init_value: 0.5
  freeze_encoder_layers: 0
  gradient_checkpointing: false
tokenizer:
  max_length: 256
  padding: max_length
  truncation: true
training:
  output_dir: ./results/roberta-base-3task_fusion
  learning_rate: 2.0e-05
  batch_size: 8
  eval_batch_size: 8
  num_epochs: 15
  weight_decay: 0.01
  warmup_steps: 20
  max_grad_norm: 1.0
  gradient_accumulation_steps: 1
  use_amp: false
  device: auto
  task_sampling: proportional
  temperature: 2.0
  early_stopping_patience: 5
  metric_for_best_model: eval_loss
  greater_is_better: false
  logging:
    service: tensorboard
    experiment_name: security_3task_fusion
    tracking_uri: null
    log_every_n_steps: 5
  checkpoint:
    save_every_n_epochs: 1
    save_best: true
    max_checkpoints: 2
  data:
    test_size: 0.2
    seed: 42
    num_workers: 0
    pin_memory: true
    pii_redaction:
      enabled: false
      strategy: mask
      detect_types: null
      hash_salt: jenga-ai-pii
      log_detections: true