# GRPO Reasoning Example
# Fine-tune a reasoning model (math, logic, step-by-step)
# Uses Group Relative Policy Optimization to select best outputs

model: TinyLlama/TinyLlama-1.1B-Chat-v1.0
data:
  path: examples/data/reasoning_math.jsonl
  format: alpaca
task: grpo
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-4
lr_scheduler_type: cosine
warmup_ratio: 0.1
weight_decay: 0.01
max_seq_length: 2048
output_dir: ./output_reasoning/
seed: 42
grpo_beta: 0.05
num_generations: 4
reward_fn: accuracy  # built-in reward function
logging_steps: 10
save_steps: 50
eval_steps: 50
eval_strategy: steps