# RLHF Step 1: Supervised Fine-Tuning (SFT)
# Pre-train the model before reward model training
# Use high-quality examples

model: TinyLlama/TinyLlama-1.1B-Chat-v1.0
data:
  path: examples/data/alpaca_tiny.jsonl
  format: alpaca
task: sft
backend: transformers
quantization: null
lora_r: 64
lora_alpha: 128
lora_dropout: 0.05
lora_target_modules:
  - q_proj
  - v_proj
batch_size: 16
gradient_accumulation_steps: 1
num_epochs: 1
learning_rate: 5e-4
lr_scheduler_type: linear
warmup_ratio: 0.05
weight_decay: 0.0
max_seq_length: 2048
output_dir: ./output_rlhf_sft/
seed: 42
logging_steps: 10
save_steps: 50
save_strategy: steps