# Default training configuration for SFT

# Learning parameters
learning_rate: 2e-5
weight_decay: 0.01
warmup_ratio: 0.1
lr_scheduler_type: linear
optim: adamw_torch

# Training parameters
num_train_epochs: 5
per_device_train_batch_size: 8
gradient_accumulation_steps: 8
bf16: true
remove_unused_columns: false

# SFT-specific parameters
max_length: 4096

# Logging and saving
report_to: tensorboard
logging_steps: 1
save_strategy: steps
save_steps: 10