Commit 284ee0a0 authored by jameskrw's avatar jameskrw Committed by YaningGao
Browse files

updated training prompts

parent a5d989ae

scripts/examples/navigation_aico/run.sh

+1 −1
Original line number Diff line number Diff line
@@ -78,5 +78,5 @@ python3 -m vagen.trainer.main_ppo \ 
    rollout_manager.n_trajectory=1 \

    rollout_manager.use_service=True \

    rollout_manager.timeout=240 \

    rollout_manager.base_url="http://localhost:5001" \

    rollout_manager.base_url="http://localhost:5000" \

    2>&1 | tee aico_navigation_vision.log
 
 No newline at end of file

scripts/exps/format/maniskill/freethink/run.sh

+7 −7
Original line number Diff line number Diff line
@@ -33,8 +33,8 @@ python3 -m vagen.trainer.main_ppo \ 
    data.train_batch_size=64 \

    data.val_batch_size=32 \

    data.max_prompt_length=1024 \

    data.max_response_length=256 \

    data.max_trajectory_length=3000 \

    data.max_response_length=150 \

    data.max_trajectory_length=4000 \

    data.image_key=images \

    data.truncation=error \

    actor_rollout_ref.model.path=Qwen/Qwen2.5-VL-3B-Instruct \
@@ -46,8 +46,8 @@ python3 -m vagen.trainer.main_ppo \ 
    actor_rollout_ref.actor.kl_loss_coef=0.001 \

    actor_rollout_ref.actor.kl_loss_type=mse \

    actor_rollout_ref.model.enable_gradient_checkpointing=True \

    actor_rollout_ref.actor.fsdp_config.param_offload=False \

    actor_rollout_ref.actor.fsdp_config.optimizer_offload=False \

    actor_rollout_ref.actor.fsdp_config.param_offload=True \

    actor_rollout_ref.actor.fsdp_config.optimizer_offload=True \

    actor_rollout_ref.rollout.log_prob_micro_batch_size_per_gpu=1 \

    actor_rollout_ref.rollout.tensor_model_parallel_size=4 \

    actor_rollout_ref.rollout.name=vllm \
@@ -77,8 +77,8 @@ python3 -m vagen.trainer.main_ppo \ 
    trainer.save_freq=90 \

    trainer.test_freq=20 \

    trainer.total_training_steps=200 \

    rollout_manager.max_turns=3 \

    rollout_manager.window_size=3 \

    rollout_manager.max_turns=4 \

    rollout_manager.window_size=4 \

    rollout_manager.use_multi_turn_reward=False \

    rollout_manager.use_loss_mask=True \

    rollout_manager.use_gae_mask=True \
@@ -87,5 +87,5 @@ python3 -m vagen.trainer.main_ppo \ 
    rollout_manager.n_trajectory=2 \

    rollout_manager.use_service=True \

    rollout_manager.timeout=240 \

    rollout_manager.base_url="http://localhost:5000" \

    rollout_manager.base_url="http://localhost:5001" \

    2>&1 | tee $EXPERIMENT_NAME.log

scripts/exps/format/navigation/freethink/env_config.yaml

+1 −1
Original line number Diff line number Diff line
@@ -5,4 +5,4 @@ env1: 
        prompt_format: free_think

        use_accuracy_reward: false

    train_size: 10000  

    test_size: 128
 
 No newline at end of file
 
    test_size: 64
 
 No newline at end of file

scripts/exps/format/navigation/freethink/run.sh

+7 −7
Original line number Diff line number Diff line
@@ -27,10 +27,10 @@ python3 -m vagen.trainer.main_ppo \ 
    data.train_files=data/$EXPERIMENT_NAME/train.parquet \

    data.val_files=data/$EXPERIMENT_NAME/test.parquet \

    data.train_batch_size=64 \

    data.val_batch_size=16 \

    data.val_batch_size=8 \

    data.max_prompt_length=1024 \

    data.max_response_length=256 \

    data.max_trajectory_length=3800 \

    data.max_response_length=150 \

    data.max_trajectory_length=4000 \

    data.image_key=images \

    data.truncation=error \

    actor_rollout_ref.model.path=Qwen/Qwen2.5-VL-3B-Instruct \
@@ -42,12 +42,12 @@ python3 -m vagen.trainer.main_ppo \ 
    actor_rollout_ref.actor.kl_loss_coef=0.001 \

    actor_rollout_ref.actor.kl_loss_type=mse \

    actor_rollout_ref.model.enable_gradient_checkpointing=True \

    actor_rollout_ref.actor.fsdp_config.param_offload=False \

    actor_rollout_ref.actor.fsdp_config.optimizer_offload=False \

    actor_rollout_ref.actor.fsdp_config.param_offload=True \

    actor_rollout_ref.actor.fsdp_config.optimizer_offload=True \

    actor_rollout_ref.rollout.log_prob_micro_batch_size_per_gpu=1 \

    actor_rollout_ref.rollout.tensor_model_parallel_size=4 \

    actor_rollout_ref.rollout.name=vllm \

    actor_rollout_ref.rollout.gpu_memory_utilization=0.3 \

    actor_rollout_ref.rollout.gpu_memory_utilization=0.1 \

    actor_rollout_ref.rollout.enable_chunked_prefill=False \

    actor_rollout_ref.rollout.enforce_eager=False \

    actor_rollout_ref.rollout.free_cache_engine=False \
@@ -83,5 +83,5 @@ python3 -m vagen.trainer.main_ppo \ 
    rollout_manager.n_trajectory=1 \

    rollout_manager.use_service=True \

    rollout_manager.timeout=240 \

    rollout_manager.base_url="http://localhost:5001" \

    rollout_manager.base_url="http://localhost:5000" \

    2>&1 | tee $EXPERIMENT_NAME.log

scripts/install.sh

+1 −1
Original line number Diff line number Diff line
@@ -32,4 +32,4 @@ echo "Installing Frozenlake dependencies" 
pip install 'gymnasium'

pip install "gymnasium[toy-text]"



echo "Installation complete, to install dependencies for other environments, refer to env/readme!"
 
echo "Installation complete, to install dependencies for other environments, refer to env/readme"

CRA Git | Maintained and supported by SUSTech CRA and CCSE