# RLHF Step 3: PPO Training
# Optimize with reinforcement learning using the reward model
# Generates outputs and ranks them with the reward model

model: TinyLlama/TinyLlama-1.1B-Chat-v1.0
data:
  path: examples/data/alpaca_tiny.jsonl
  format: alpaca
task: ppo
backend: transformers
quantization: null
lora_r: 64
lora_alpha: 128
lora_dropout: 0.05
lora_target_modules:
  - q_proj
  - v_proj
batch_size: 8
gradient_accumulation_steps: 2
num_epochs: 2
learning_rate: 1e-5
lr_scheduler_type: cosine
warmup_ratio: 0.1
weight_decay: 0.01
max_seq_length: 2048
output_dir: ./output_ppo/
seed: 42
ppo_epochs: 4
ppo_clip_ratio: 0.2
ppo_kl_penalty: 0.05
reward_model: ./output_reward_model/
logging_steps: 10
save_steps: 50