conf.yaml

common_args:
  template_name: "llm_finetuning"
  project_name: "FLockLLM_finetune"

data_args:
  data_path: "/dataset.json"

model_args:
  foundation_model_name: "meta-llama/Meta-Llama-3.1-8B," # meta-llama/Meta-Llama-3.1-8B, Qwen/Qwen2-7B, google/gemma-2-2b, mistralai/Mistral-7B-v0.1,
  foundation_model_source: "huggingface" # Options: "huggingface", "flock_s3" Defaults: "huggingface"
  finetune_adapter: "qlora"
  lora_dropout: 0.05
  lora_target_modules: [] # Options: "q_proj","k_proj","v_proj","o_proj" Defaults: [] (let system auto search)

train_args:
  proposer_train_batch_size: 32
  proposer_train_micro_batch_size: 8
  proposer_num_epochs: 1
  proposer_learning_rate: 0.0003
  proposer_val_set_size: 16
  proposer_save_steps: 3
  proposer_train_group_by_length: false
  proposer_train_optimizer_name: "paged_adamw_8bit"
  proposer_train_lr_scheduler_type: "constant"
  proposer_train_warmup_steps: 1
  proposer_train_weight_decay: 0.05
  block_size: 8
  cutoff_len: 1024
  federated_optimizer_name: "fedavg"
  device_map: "cuda:0"

evaluation_args:
  voter_val_set_size: 5

tracking_args:
  out_put_root: "output"
  finetune_adapter_checkpoint_save_dir: "checkpoints/llama-3.1-8b"
  proposer_train_gradient_checkpointing: true
  proposer_train_logging_steps: 10
  proposer_model_save_steps: 40
  save_total_limit: 3

nvflare_args:
  # Notice: For everytime start a simulation, the simulator will wipe eveything in the workspace
  nvflare_workspace: "/tmp/nvflare/jobs/workdir"
  nvflare_executor_name: "flockit"
  nvflare_output_dir: "/tmp/nvflare/"