qwen2_5_vl_7b_geo3k_reinforce.sh

set -x

MODEL_PATH=Qwen/Qwen2.5-VL-7B-Instruct  # replace it with your local file path

FORMAT_PROMPT="""You FIRST think about the reasoning process as an internal monologue and then provide the final answer.
 The reasoning process MUST BE enclosed within <think> </think> tags. The final answer MUST BE put in \boxed{}."""

python3 -m verl.trainer.main \
    config=examples/config.yaml \
    data.train_files=hiyouga/geometry3k@train \
    data.val_files=hiyouga/geometry3k@test \
    data.format_prompt="${FORMAT_PROMPT}" \
    worker.actor.model.model_path=${MODEL_PATH} \
    algorithm.adv_estimator=reinforce_plus_plus \
    algorithm.use_kl_loss=false \
    algorithm.kl_penalty=kl \
    algorithm.kl_coef=1.0e-3 \
    trainer.experiment_name=qwen2_5_vl_7b_geo_reinforce_pp \
    trainer.n_gpus_per_node=8