data_cfgs:
  eval_data_files: {}
  eval_datasets: {}
  eval_optional_args: []
  eval_size: {}
  eval_split: {}
  eval_subset: {}
  eval_template: {}
  train_data_files: pre_tokenized/train_no_pixel_values.pt
  train_datasets: /aifs4su/yaodong/hantao/datasets/MMInstruct-GPT4V_mistral-7b_cosi_cut/merged//top1-100_valid
  train_name: {}
  train_optional_args: []
  train_size: {}
  train_split: {}
  train_template: {}
logger_cfgs:
  cache_dir: {}
  log_project: align-anything
  log_run_name: sft
  log_type: wandb
  output_dir: ../outputs/chameleon_sft/top1-100_valid
  save_total_limit: 12
model_cfgs:
  model_max_length: 4096
  model_name_or_path: /aifs4su/yaodong/hantao/models/chameleon-7b-hf
  trust_remote_code: true
special_tokens: {}
train_cfgs:
  adam_betas:
  - 0.9
  - 0.95
  adam_epsilon: 1.0e-08
  bf16: true
  ds_cfgs: ds_z3_config.json
  epochs: 3
  eval_interval: 1000
  eval_strategy: steps
  fp16: false
  freeze_language_model: false
  gradient_accumulation_steps: 16
  gradient_checkpointing: true
  learning_rate: 2.0e-05
  load_checkpoint: false
  lr_scheduler_type: cosine
  lr_warmup_ratio: 0.03
  max_grad_norm: 1.0
  per_device_eval_batch_size: 1
  per_device_train_batch_size: 1
  save_checkpoint: true
  seed: 42
  weight_decay: 0.0