sft.yaml

# SFT (Supervised Fine-Tuning) Configuration Example

method: sft

# Data configuration
data:
  train_files: ./output/learnable.jsonl
  val_files: ./output/val.jsonl
  train_batch_size: 32
  micro_batch_size_per_gpu: 4
  max_length: 2048                 # Maximum sequence length (prompt + response)
  truncation: right                # Options: left, right, error
  prompt_key: input                # Key for prompt field in JSONL
  response_key: output             # Key for response field in JSONL

# Model configuration
model:
  path: Qwen/Qwen2.5-7B            # HuggingFace model ID or local path
  enable_gradient_checkpointing: true
  use_remove_padding: true         # Removes padding tokens to save memory
  use_liger: false                 # Use Liger kernel for faster training

  # LoRA configuration (Parameter-Efficient Fine-Tuning)
  # Set lora_rank > 0 to enable LoRA
  lora_rank: 0                     # 0 = full fine-tuning, 32/64 = LoRA enabled
  lora_alpha: 16                   # LoRA scaling factor
  target_modules: all-linear       # Modules to apply LoRA: all-linear, q_proj,v_proj,k_proj,o_proj

# Optimizer configuration
optim:
  lr: 1e-5                         # Learning rate

# Parallelism settings
ulysses_sequence_parallel_size: 1  # Sequence parallelism for long sequences

# Wandb configuration (optional, for experiment tracking)
wandb:
  api_key: your-api-key            # Or set WANDB_API_KEY environment variable
  entity:                          # Wandb username or team (auto-detected if not provided)

# Trainer configuration
trainer:
  project_name: sdg-sft
  experiment_name: sft-qwen2.5-7b
  logger: console,wandb            # Options: console, wandb, tensorboard (comma-separated)
  total_epochs: 1                  # Training epochs
  save_freq: 100                   # Save checkpoint every N steps
  nnodes: 1                        # Number of nodes
  n_gpus_per_node: 1               # GPUs per node
  default_local_dir: ./checkpoints/sft

  # Wandb proxy (optional, for environments requiring proxy)
  # wandb_proxy: http://proxy.example.com:8080