config.yaml

# Giant-Killer NLP Configuration
# Dendritic Optimization Hackathon

# Model Configuration
model:
  name: "prajjwal1/bert-tiny"  # 2 layers, 128 hidden size, ~4M params
  num_labels: 2  # Binary classification (toxic/non-toxic)
  hidden_dropout_prob: 0.1
  attention_probs_dropout_prob: 0.1

# Data Configuration
data:
  dataset_name: "jigsaw_toxicity_pred"  # Jigsaw Unintended Bias dataset
  max_length: 128  # For fast real-time processing
  train_split: "train"
  val_split: "validation"
  test_split: "test"
  batch_size: 32
  num_workers: 4
  
# Training Configuration
training:
  epochs: 10
  learning_rate: 2.0e-5
  weight_decay: 0.01
  warmup_steps: 500
  max_grad_norm: 1.0
  
  # Scheduler configuration (handled by PAI tracker)
  scheduler:
    step_size: 1
    gamma: 0.1
    
  # Early stopping
  early_stopping:
    patience: 3
    min_delta: 0.001

# Perforated AI Configuration
perforated_ai:
  enabled: true
  # Dendrite learning starts after base model plateaus
  dendrite_learning:
    enabled: true
    correlation_threshold: 0.95

# Evaluation Configuration
evaluation:
  metrics:
    - accuracy
    - f1
    - precision
    - recall
    - auc_roc
  
  # Benchmarking configuration
  benchmark:
    num_samples: 100
    device: "cpu"  # Benchmark on CPU for edge deployment
    warm_up_runs: 10

# Quantization Configuration
quantization:
  enabled: true
  dtype: "qint8"
  # Layers to quantize
  layers:
    - torch.nn.Linear
    - torch.nn.Embedding

# Logging Configuration
logging:
  level: "INFO"
  tensorboard: true
  log_dir: "logs/"
  save_dir: "checkpoints/"
  save_every_n_epochs: 1

# Reproducibility
seed: 42
deterministic: true