training_args.json

{
  "lora_r": 16,
  "lora_alpha": 16,
  "lora_dropout": 0.05,
  "lora_target_modules": "q_proj,v_proj",
  "tuning_strategy": "lora",
  "num_trainable_layers": 2,
  "output_dir": "sft_prefill/prompt_id_3/qwen2-VL-7B-Instruct-syn-count-lora",
  "num_train_epochs": 2,
  "learning_rate": 0.0002,
  "per_device_train_batch_size": 16,
  "per_device_eval_batch_size": 16,
  "gradient_accumulation_steps": 1,
  "logging_steps": 10,
  "eval_steps": 200,
  "save_steps": 200,
  "warmup_ratio": 0.03,
  "weight_decay": 0.0,
  "max_grad_norm": 0.3,
  "lr_scheduler_type": "constant",
  "bf16": true,
  "tf32": true,
  "gradient_checkpointing": true,
  "optim": "adamw_torch_fused",
  "ft_type": "SFT_Prefill",
  "data_type": "small",
  "prompt_id": 3
}