configs/model_config.yaml

# Model Configuration
model:
  # Base models
  gemma_model_name: "google/gemma-3-270m"  # 270M parameter model - fast training on A100
  vision_model_name: "openai/clip-vit-large-patch14"

  # Model settings - vision-language only
  projector_hidden_dim: 2048  # Larger projection for better alignment
  
  # LoRA configuration - optimized for multimodal
  lora:
    r: 64          # Higher rank for better multimodal understanding
    alpha: 128     # Higher alpha for better learning
    dropout: 0.1   # Slightly higher dropout for regularization
    target_modules:
      - "q_proj"
      - "v_proj"
      - "k_proj"
      - "o_proj"
      - "gate_proj"
      - "up_proj"
      - "down_proj"
  
  # Quantization
  use_4bit: true
  bnb_4bit_compute_dtype: "bfloat16"
  bnb_4bit_quant_type: "nf4"
  use_nested_quant: false

# Tokenizer settings
tokenizer:
  padding_side: "right"
  truncation: true
  max_length: 512
  add_special_tokens: true

# Special tokens
special_tokens:
  image_token: "<image>"
  pad_token: "<pad>"