Clarify differences between examples/rl/grpo GSM8K recipes and scripts/grpo_demo_llama3_qwen2.py

Hi, thanks for open-sourcing Tunix!

I’m a bit confused about the two GRPO entry points that seem to cover similar use cases:

Shell scripts under examples/rl/grpo/gsm8k (e.g. run_llama3.2_1b.sh, run_gemma3_1b.sh), which call python -m tunix.cli.grpo_main ... with a base_config.yaml, dataset_name="gsm8k", reward_functions="['tunix/cli/reward_fn/gsm8k.py']", and various overrides for steps, LR, checkpointing, etc.

The Python demo scripts/grpo_demo_llama3_qwen2.py, which sets up GRPO for Llama 3 / Qwen 2 with its own argument parser (e.g. --model-version) .

So what is the main difference? When should I choose this one instead of the other?