OGER: A Robust Offline-Guided Exploration Reward for Hybrid Reinforcement Learning

📖Introduction

OGER is a framework that enhances reasoning capabilities by unifying offline teacher guidance and online RL through a specialized reward modeling lens. Specifically, our framework leverages multi-source teacher trajectories for collaborative training to solidify foundational reasoning capabilities. Concurrently, we construct a divergence-based exploration reward to quantify the semantic disparity between online and offline trajectories, facilitating a profound synergy between expert imitation and autonomous discovery. To ensure training stability, we implement a strategic hybrid sampling mechanism that integrates offline expert data directly into the online training batches. Furthermore, we refine this exploration signal by incorporating the policy model's token-level entropy distribution, enabling fine-grained control to incentivize novel reasoning behaviors and mitigate the risk of premature convergence.

Key Highlights:

• Multi-source Offline Trajectories
• Offline-Guided Exploration Reward
• OGER Reward with Hybrid Set

---

✨Getting Started

Installation

You can install OGER dependencies by running the following commands:

conda create -n oger python=3.10
conda activate oger
cd oger
pip install -r requirements.txt
pip install -e .
cd verl
pip install -e .

Quick Start

We provide an example script to train OGER on our multi-source offline subset of OpenR1-Math-220k. You can run the following command to train OGER:

conda activate oger

cd ./verl

save_dir=/path/to/your/dir

export MODEL_PATH=/path/to/your/model
export PROJECT_NAME=project_name
export EXPERIMENT_NAME=oger
export TENSORBOARD_DIR=/path/to/your/log


# Set XFormers backend to avoid CUDA errors
export VLLM_ATTENTION_BACKEND=XFORMERS
export DATA_DIR=/path/to/your/data

# Train over a single node, 8 AH200 GPUs.
python3 -m verl.mix_src.main_mix_ppo \
    algorithm.adv_estimator=grpo \
    data.train_files=$DATA_DIR/luffy_offline_samples.parquet \
    data.val_files=$DATA_DIR/valid.parquet \
    data.train_batch_size=128 \
    data.val_batch_size=512 \
    data.max_prompt_length=1024 \
    data.max_response_length=8192 \
    actor_rollout_ref.model.path=$MODEL_PATH \
    actor_rollout_ref.actor.optim.lr=1e-6 \
    actor_rollout_ref.model.use_remove_padding=True \
    actor_rollout_ref.actor.ppo_mini_batch_size=64 \
    actor_rollout_ref.actor.ppo_micro_batch_size=64 \
    actor_rollout_ref.actor.use_dynamic_bsz=True \
    actor_rollout_ref.actor.ppo_max_token_len_per_gpu=32768 \
    actor_rollout_ref.actor.kl_loss_coef=0.00 \
    actor_rollout_ref.actor.kl_loss_type=low_var_kl \
    actor_rollout_ref.actor.ulysses_sequence_parallel_size=1 \
    actor_rollout_ref.model.enable_gradient_checkpointing=True \
    actor_rollout_ref.actor.fsdp_config.param_offload=False \
    actor_rollout_ref.actor.fsdp_config.grad_offload=False \
    actor_rollout_ref.actor.fsdp_config.optimizer_offload=False \
    actor_rollout_ref.rollout.tensor_model_parallel_size=2 \
    actor_rollout_ref.rollout.name=vllm \
    actor_rollout_ref.rollout.temperature=1.0 \
    actor_rollout_ref.rollout.val_temperature=0.6 \
    actor_rollout_ref.rollout.gpu_memory_utilization=0.60 \
    actor_rollout_ref.rollout.n=8 \
    actor_rollout_ref.rollout.n_val=1 \
    actor_rollout_ref.ref.fsdp_config.param_offload=True \
    actor_rollout_ref.rollout.max_prefix_len=8192 \
    algorithm.kl_ctrl.kl_coef=0.000 \
    actor_rollout_ref.actor.entropy_coeff=0.001 \
    trainer.critic_warmup=0 \
    trainer.logger=['console','tensorboard'] \
    trainer.project_name="$PROJECT_NAME" \
    trainer.experiment_name="$EXPERIMENT_NAME" \
    +trainer.val_before_train=False \
    +trainer.max_replace_num=1 \
    +trainer.similarity_embedding=True \
    +trainer.similarity_embedding_model=/path/to/embedding_model \
    +trainer.replace_mode="similarity_mean_des" \
    +trainer.use_similarity_score=True \
    +trainer.similarity_shaping=False \
    +trainer.similarity_reverse=True \
    +trainer.similarity_reverse_fail=False \
    +trainer.last_token_enp_reward='exp_neg_enp' \
    +trainer.skip_all_success=False \
    +trainer.use_history_samples=False \
    +trainer.use_mu_schedule=False \
    +trainer.repeating_check=False \
    +trainer.similarity_reranker=False \
    trainer.n_gpus_per_node=8 \
    trainer.nnodes=1 \
    trainer.save_freq=50 \
    trainer.test_freq=10 \
    +actor_rollout_ref.rollout.replace_after_rm=True \
    trainer.default_local_dir=$save_dir/$PROJECT_NAME/$EXPERIMENT_NAME \
    actor_rollout_ref.actor.use_kl_loss=False \
    actor_rollout_ref.actor.use_sft_prefix_reward=False \
    actor_rollout_ref.rollout.prefix_share_across_samples=False \
    actor_rollout_ref.rollout.prefix_strategy=random \
    actor_rollout_ref.rollout.n_prefix=1 \
    actor_rollout_ref.rollout.min_prefix_ratio=0.0 \
    actor_rollout_ref.rollout.max_prefix_ratio=0.0 \
    actor_rollout_ref.rollout.prefix_reward_weight_alpha=1.0 \
    actor_rollout_ref.ref.use_ref=False \
    actor_rollout_ref.actor.use_off_policy_loss=True \
    actor_rollout_ref.actor.off_policy_normalize=False \
    actor_rollout_ref.actor.off_policy_reshape="p_div_p_0.1" \
    actor_rollout_ref.actor.off_policy_loss_impl=token \
    algorithm.grpo_use_std=False \
    actor_rollout_ref.actor.loss_remove_token_mean=True \
    actor_rollout_ref.actor.loss_remove_clip=True \
    data.reward_impl_version=3 \
    data.shuffle=True \
    trainer.default_hdfs_dir=null \
    trainer.max_optim_to_keep=-1 \
    trainer.total_training_steps=2000 "${@:1}"