configs/default.yaml

# KerdosAI Default Configuration

# Model Configuration
base_model: "gpt2"  # Base model name or path
model_revision: null  # Specific model revision/commit
trust_remote_code: false  # Whether to trust remote code
device: null  # Device to use (cuda/cpu), null for auto-detection

# LoRA Configuration
lora:
  enabled: true
  r: 8  # LoRA rank
  alpha: 32  # LoRA alpha
  dropout: 0.1  # LoRA dropout
  target_modules: null  # Auto-detect if null

# Quantization Configuration
quantization:
  enabled: false
  bits: 4  # 4 or 8
  use_double_quant: true
  quant_type: "nf4"  # nf4 or fp4
  compute_dtype: "float16"  # float16, bfloat16, or float32

# Training Configuration
training:
  epochs: 3
  batch_size: 4
  learning_rate: 0.00002  # 2e-5
  warmup_steps: 100
  gradient_accumulation_steps: 1
  max_grad_norm: 1.0
  weight_decay: 0.01
  logging_steps: 10
  save_steps: 100
  eval_steps: 100
  max_seq_length: 512
  seed: 42
  fp16: false
  bf16: false

# Data Configuration
data:
  train_file: null  # Path to training data
  validation_file: null  # Path to validation data
  test_file: null  # Path to test data
  dataset_name: null  # HuggingFace dataset name
  dataset_config: null  # Dataset configuration
  text_column: "text"  # Column name for text data
  max_samples: null  # Limit number of samples (null for all)
  preprocessing_num_workers: 4

# Deployment Configuration
deployment:
  type: "rest"  # rest, docker, or kubernetes
  host: "0.0.0.0"
  port: 8000
  workers: 1
  max_batch_size: 8
  timeout: 60

# Monitoring Configuration
monitoring:
  enabled: true
  wandb_project: null  # W&B project name
  wandb_entity: null  # W&B entity/team name
  tensorboard_dir: "./runs"
  log_model: false

# Output Configuration
output_dir: "./output"
checkpoint_dir: "./checkpoints"
cache_dir: null  # HuggingFace cache directory