data:
  train_files: ""
  val_files: ""
  prompt_key: problem
  answer_key: answer
  image_key: images
  video_key: videos
  image_dir: null
  video_fps: 2.0
  max_prompt_length: 16384
  max_response_length: 4096
  rollout_batch_size: 128
  mini_rollout_batch_size: null
  val_batch_size: 1024
  format_prompt: ""
  override_chat_template: null
  shuffle: true
  seed: 1
  min_pixels: 3136
  max_pixels: 1048576
  resize_size: 336
  filter_overlong_prompts: false

algorithm:
  adv_estimator: grpo
  disable_kl: true
  use_kl_loss: true
  kl_penalty: low_var_kl
  kl_coef: 1.0e-2
  online_filtering: false
  filter_key: accuracy
  filter_low: 0.0
  filter_high: 1.0

worker:
  actor:
    global_batch_size: 32
    micro_batch_size_per_device_for_update: 1
    micro_batch_size_per_device_for_experience: 1
    max_grad_norm: 1.0
    padding_free: true
    dynamic_batching: true
    ulysses_size: 1
    model:
      model_path: ""
      enable_gradient_checkpointing: true
      trust_remote_code: false
      freeze_vision_tower: true
    optim:
      lr: 5.0e-6
      weight_decay: 1.0e-2
      strategy: adamw
      lr_warmup_ratio: 0.0
    fsdp:
      enable_full_shard: true
      enable_cpu_offload: false
      enable_rank0_init: true
    offload:
      offload_params: false
      offload_optimizer: false

  rollout:
    n: 8
    temperature: 1.0
    top_p: 1.0
    limit_images: 0
    gpu_memory_utilization: 0.7
    enforce_eager: false
    enable_chunked_prefill: false
    tensor_parallel_size: 4
    disable_tqdm: true
    max_num_batched_tokens: 20480
    val_override_config:
      temperature: 0.7
      top_p: 0.95
      n: 1

  ref:
    fsdp:
      enable_full_shard: true
      enable_cpu_offload: false
      enable_rank0_init: true
    offload:
      offload_params: false

  reward:
    reward_type: batch
    reward_function: EasyR1/verl/reward_function/onethinker_reward.py:compute_score

trainer:
  total_epochs: 1
  max_steps: null
  project_name: easy_r1
  experiment_name: ""
  logger: ["file", "wandb"]
  nnodes: 1
  n_gpus_per_node: 2
  max_try_make_batch: 20
  val_freq: -1
  val_before_train: false
  val_only: false
  val_generations_to_log: 3
  save_freq: 1
  save_limit: 100000
  save_model_only: false
  save_checkpoint_path: ""
  load_checkpoint_path: null
  find_last_checkpoint: false