callbacks:
  save_plan: true
  preview_samples: 2
dataset:
  dataset_name: nwpu
  dataset_root: ${path.data_root}/NWPU-RESISC45
  train_split: train
  eval_split: val
  max_train_samples: null
  max_eval_samples: null
logger:
  name: console
  level: INFO
model:
  model_name: LFM2.5-VL-450M
  base_model: LiquidAI/LFM2.5-VL-450M
  output_dir: /tmp/terra-q-base-eval
  image_size: 512
  adapter:
    enabled: true
    method: lora
    rank: 8
    alpha: 16
    dropout: 0.0
    target_modules:
    - q_proj
    - k_proj
    - v_proj
    - o_proj
    - gate_proj
    - up_proj
    - down_proj
path:
  data_root: ${hydra:runtime.cwd}/data
  output_root: outputs
  work_dir: ${path.output_root}/${experiment_name}
trainer:
  epochs: 3
  learning_rate: 5.0e-06
  per_device_batch_size: 16
  gradient_accumulation_steps: 2
  max_new_tokens: 512
  temperature: 0.1
  seed: 42
  warmup_steps: 100
  max_seq_length: 2048
  report_to: none
  dataloader_num_workers: 4
  eval_steps: 400
  lr_scheduler_type: cosine_with_min_lr
  lr_scheduler_min_lr: 5.0e-07
  weight_decay: 0.0001
  max_grad_norm: 1.0
  early_stopping_patience: 2
experiment_name: lfm25-vl-450m_nwpu_lora-r8-a16_ep3-lr5e-6-min5e-7-w100-wd1e-4