data:
  tokenizer: null
  train_files: dapo_ds_train_sample.parquet
  val_files: matheval.parquet
  prompt_key: prompt
  reward_fn_key: data_source
  max_prompt_length: 1024
  max_response_length: 2048
  train_batch_size: 128
  val_batch_size: 640
  return_raw_input_ids: false
  return_raw_chat: false
  shuffle: true
  filter_overlong_prompts: false
  filter_overlong_prompts_workers: 1
  truncation: left
  image_key: images
  custom_cls:
    path: null
    name: null
actor_rollout_ref:
  hybrid_engine: true
  model:
    path: Qwen/Qwen2.5-Math-7B
    external_lib: null
    override_config: {}
    enable_gradient_checkpointing: true
    use_remove_padding: false
    use_liger: false
    save_hf_repo_id: RyanYr/brm-dapo-qwen2.5math-7B-base_lr2e-6-beta0.002
    tokenizer_chat_template: null
  actor:
    brm:
      norm_factor: value
      value_constant: null
    buffer:
      buffer_type: null
      offline_dataset_buffer:
        train_files: null
        response_key: response
        response_truncation: right
        shuffle: true
        update_size: 128
    strategy: fsdp
    actor_micro_batch_size: null
    actor_micro_batch_size_per_gpu: 2
    use_dynamic_bsz: false
    actor_max_token_len_per_gpu: 16384
    grad_clip: 1.0
    use_torch_compile: true
    shuffle: false
    ulysses_sequence_parallel_size: 1
    checkpoint:
      contents:
      - model
      - optimizer
      - extra
    optim:
      lr: 2.0e-06
      lr_warmup_steps: -1
      lr_warmup_steps_ratio: 0.0
      min_lr_ratio: null
      warmup_style: constant
      total_training_steps: 800
      weight_decay: 0.01
    fsdp_config:
      wrap_policy:
        min_num_params: 0
      param_offload: false
      optimizer_offload: false
      fsdp_size: -1
    report_entropy: true
    actor_mini_batch_size: 128
    use_kl_loss: false
    actor_epochs: 1
  ref:
    ref_model_path: Qwen/Qwen2.5-Math-7B
    strategy: fsdp
    fsdp_config:
      param_offload: false
      wrap_policy:
        min_num_params: 0
    log_prob_micro_batch_size: null
    log_prob_micro_batch_size_per_gpu: 4
    log_prob_use_dynamic_bsz: false
    log_prob_max_token_len_per_gpu: 16384
    ulysses_sequence_parallel_size: 1
  rollout:
    name: vllm
    temperature: 1.0
    top_k: -1
    top_p: 1
    use_fire_sampling: false
    prompt_length: 1024
    response_length: 2048
    dtype: bfloat16
    gpu_memory_utilization: 0.7
    ignore_eos: false
    enforce_eager: false
    free_cache_engine: false
    load_format: dummy_dtensor
    tensor_model_parallel_size: 4
    max_num_batched_tokens: 4072
    max_model_len: null
    max_num_seqs: 1024
    log_prob_micro_batch_size: null
    log_prob_micro_batch_size_per_gpu: 4
    log_prob_use_dynamic_bsz: false
    log_prob_max_token_len_per_gpu: 16384
    disable_log_stats: true
    enable_chunked_prefill: true
    do_sample: true
    'n': 1
    engine_kwargs:
      swap_space: null
    val_kwargs:
      top_k: -1
      top_p: 1.0
      temperature: 0.6
      'n': 32
      do_sample: true
reward_model:
  enable: false
  strategy: fsdp
  model:
    input_tokenizer: Qwen/Qwen2.5-Math-7B
    path: ~/models/FsfairX-LLaMA3-RM-v0.1
    external_lib: null
    use_remove_padding: false
    fsdp_config:
      wrap_policy:
        min_num_params: 0
      param_offload: false
      fsdp_size: -1
  actor_micro_batch_size: null
  actor_micro_batch_size_per_gpu: null
  max_length: null
  ulysses_sequence_parallel_size: 1
  use_dynamic_bsz: false
  forward_max_token_len_per_gpu: 16384
  reward_manager: prime
  reward_kwargs:
    format_reward: 0.0
    format_type: null
custom_reward_function:
  path: null
  name: compute_score
trainer:
  balance_batch: true
  total_epochs: 50
  total_training_steps: 800
  project_name: value-LLM
  experiment_name: brm-dapo-qwen2.5math-7B-base_lr2e-6-beta0.002
  logger:
  - console
  - wandb
  log_val_generations: 0
  nnodes: 1
  n_gpus_per_node: 4
  save_freq: 40
  resume_mode: auto
  resume_from_path: null
  val_before_train: false
  test_freq: -1
  default_hdfs_dir: null
  del_local_ckpt_after_load: false
  default_local_dir: ./BRM
  max_actor_ckpt_to_keep: 1
  ray_wait_register_center_timeout: 300
  hf_token: null
  resume_from_hf:
    enable: false
    hf_repo_id: null
    hf_token: null
    revision: main
algorithm:
  kl_ctrl:
    kl_coef: 0.002
  use_kl_in_reward: false
  adv_estimator: none