colpali-main/scripts/configs/text_only/train_colllama_model_debug.yaml

config:
  (): colpali_engine.utils.train_colpali_engine_models.ColModelTrainingConfig
  model:
    (): colpali_engine.utils.wrapper.AutoColModelWrapper
    pretrained_model_name_or_path: "HuggingFaceM4/tiny-random-LlamaForCausalLM"
    # attn_implementation: "eager"
    quantization_config:
      (): transformers.BitsAndBytesConfig
      load_in_4bit: true
      bnb_4bit_quant_type: "nf4"
      bnb_4bit_compute_dtype:  "float16"
      bnb_4bit_use_double_quant: true

  dataset_loading_func: !ext colpali_engine.utils.dataset_transformation.load_manu_embeddings
  loss_func:
    (): colpali_engine.loss.colbert_loss.ColbertLoss
  max_length: 256
  run_eval: true
  add_suffix: true
  tr_args:
    (): transformers.training_args.TrainingArguments
    output_dir: null
    overwrite_output_dir: true
    num_train_epochs: 1
    per_device_train_batch_size: 64
    gradient_accumulation_steps: 2
    per_device_eval_batch_size: 8
    dataloader_num_workers: 8
    bf16: false
    save_steps: 500
    logging_steps: 10
    warmup_steps: 500
    learning_rate: 5e-5
    save_total_limit: 1
    optim: "paged_adamw_8bit"

  peft_config:
    (): peft.LoraConfig
    r: 16
    lora_alpha: 32
    lora_dropout: 0.05
    bias: "none"
    task_type: "FEATURE_EXTRACTION"
    target_modules:
      - 'up_proj'
      - 'down_proj'
      - 'gate_proj'
      - 'k_proj'
      - 'q_proj'
      - 'v_proj'
      - 'o_proj'