Amu/t1-3B-grpo

Train

Eval

qwen-2.5-3B-Instruct

task_type	metric	dataset_name	subset_name	average_score	count	average_score
math	AveragePass@1	math_500	default	0.648	500
math	AveragePass@1	gpqa	gpqa_diamond	0.2677	198	0.5192
math	AveragePass@1	aime24	default	0.0333	30

t1-3B

task_type	metric	dataset_name	subset_name	average_score	count	average_score
math	AveragePass@1	math_500	default	0.698	500
math	AveragePass@1	gpqa	gpqa_diamond	0.3182	198	0.5714
math	AveragePass@1	aime24	default	0.1333	30

t1-3B-grpo

task_type	metric	dataset_name	subset_name	average_score	count	average_score
math	AveragePass@1	math_500	default	0.77	500
math	AveragePass@1	gpqa	gpqa_diamond	0.2879	198	0.6113
math	AveragePass@1	aime24	default	0.1	30

Reproduce

I use t1-3b as a base model which it is trained by using t1-101k.

I use the grpo algorithm in deepscaler.

the below is the grpo script:

• t1-3b-grpo-8k

CUDA_VISIBLE_DEVICES=2,3,4,5,6,7 python3 -m verl.trainer.main_ppo \
    algorithm.adv_estimator=grpo \
    data.train_files=/root/deepscaler/data/train.parquet \
    data.val_files=/root/deepscaler/data/aime.parquet \
    data.train_batch_size=96 \
    data.val_batch_size=288 \
    data.max_prompt_length=1024 \
    data.max_response_length=8192 \
    actor_rollout_ref.model.path=/workspace/R1/model/t1-3B  \
    actor_rollout_ref.actor.optim.lr=1e-6 \
    actor_rollout_ref.model.use_remove_padding=True \
    actor_rollout_ref.actor.ppo_mini_batch_size=48 \
    actor_rollout_ref.actor.ppo_micro_batch_size=48 \
    actor_rollout_ref.actor.use_dynamic_bsz=True \
    actor_rollout_ref.actor.ppo_max_token_len_per_gpu=32768 \
    actor_rollout_ref.actor.use_kl_loss=True \
    actor_rollout_ref.actor.kl_loss_coef=0.001 \
    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=1 \
    actor_rollout_ref.rollout.name=vllm \
    actor_rollout_ref.rollout.temperature=0.6 \
    actor_rollout_ref.rollout.val_temperature=0.6 \
    actor_rollout_ref.rollout.gpu_memory_utilization=0.85 \
    actor_rollout_ref.rollout.n=12 \
    actor_rollout_ref.rollout.n_val=6 \
    actor_rollout_ref.ref.fsdp_config.param_offload=True \
    algorithm.kl_ctrl.kl_coef=0.001 \
    trainer.critic_warmup=0 \
    trainer.logger=['console','wandb'] \
    trainer.project_name='t1-3b' \
    trainer.experiment_name='t1-3b-grpo-8k' \
    +trainer.val_before_train=True \
    trainer.n_gpus_per_node=6 \
    trainer.nnodes=1 \
    trainer.save_freq=50 \
    trainer.test_freq=50 \
    trainer.default_hdfs_dir=null \
    trainer.total_epochs=5

• t1-3b-grpo-16k

CUDA_VISIBLE_DEVICES=2,3,4,5,6,7 python3 -m verl.trainer.main_ppo \
    algorithm.adv_estimator=grpo \
    data.train_files=/root/deepscaler/data/train.parquet \
    data.val_files=/root/deepscaler/data/aime.parquet \
    data.train_batch_size=48 \
    data.val_batch_size=144 \
    data.max_prompt_length=1024 \
    data.max_response_length=16384 \
    actor_rollout_ref.model.path=/workspace/R1/deepscaler/checkpoints/t1-3b/t1-3b-grpo-8k/actor/global_step_450  \
    actor_rollout_ref.actor.optim.lr=1e-6 \
    actor_rollout_ref.model.use_remove_padding=True \
    actor_rollout_ref.actor.ppo_mini_batch_size=48 \
    actor_rollout_ref.actor.ppo_micro_batch_size=48 \
    actor_rollout_ref.actor.use_dynamic_bsz=True \
    actor_rollout_ref.actor.ppo_max_token_len_per_gpu=32768 \
    actor_rollout_ref.actor.use_kl_loss=True \
    actor_rollout_ref.actor.kl_loss_coef=0.001 \
    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=1 \
    actor_rollout_ref.rollout.name=vllm \
    actor_rollout_ref.rollout.temperature=0.6 \
    actor_rollout_ref.rollout.val_temperature=0.6 \
    actor_rollout_ref.rollout.gpu_memory_utilization=0.85 \
    actor_rollout_ref.rollout.n=12 \
    actor_rollout_ref.rollout.n_val=6 \
    actor_rollout_ref.ref.fsdp_config.param_offload=True \
    algorithm.kl_ctrl.kl_coef=0.001 \
    trainer.critic_warmup=0 \
    trainer.logger=['console','wandb'] \
    trainer.project_name='t1-3b' \
    trainer.experiment_name='t1-3b-grpo-16k' \
    +trainer.val_before_train=True \
    trainer.n_gpus_per_node=6 \
    trainer.nnodes=1 \
    trainer.save_freq=20 \
    trainer.test_freq=20 \
    trainer.default_hdfs_dir=null \
    trainer.total_epochs=5

• t1-3b-grpo-32k

CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 python3 -m verl.trainer.main_ppo \
    algorithm.adv_estimator=grpo \
    data.train_files=/root/deepscaler/data/train.parquet \
    data.val_files=/root/deepscaler/data/aime.parquet \
    data.train_batch_size=16 \
    data.val_batch_size=16 \
    data.max_prompt_length=1024 \
    data.max_response_length=31744 \
    actor_rollout_ref.model.path=/workspace/R1/deepscaler/checkpoints/t1-3b/t1-3b-grpo-16k/actor/global_step_460  \
    actor_rollout_ref.actor.optim.lr=1e-6 \
    actor_rollout_ref.model.use_remove_padding=True \
    actor_rollout_ref.actor.ppo_mini_batch_size=16 \
    actor_rollout_ref.actor.ppo_micro_batch_size=16 \
    actor_rollout_ref.actor.use_dynamic_bsz=True \
    actor_rollout_ref.actor.ppo_max_token_len_per_gpu=32768 \
    actor_rollout_ref.actor.use_kl_loss=True \
    actor_rollout_ref.actor.kl_loss_coef=0.001 \
    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=1 \
    actor_rollout_ref.rollout.name=vllm \
    actor_rollout_ref.rollout.temperature=0.6 \
    actor_rollout_ref.rollout.val_temperature=0.6 \
    actor_rollout_ref.rollout.gpu_memory_utilization=0.6 \
    actor_rollout_ref.rollout.n=8 \
    actor_rollout_ref.rollout.n_val=8 \
    actor_rollout_ref.ref.fsdp_config.param_offload=True \
    algorithm.kl_ctrl.kl_coef=0.001 \
    trainer.critic_warmup=0 \
    trainer.logger=['console','wandb'] \
    trainer.project_name='t1-3b' \
    trainer.experiment_name='t1-3b-grpo-32k' \
    +trainer.val_before_train=True \
    trainer.n_gpus_per_node=8 \
    trainer.nnodes=1 \
    trainer.save_freq=20 \
    trainer.test_freq=20 \
    trainer.default_hdfs_dir=null \
    trainer.total_epochs=5

What's the next

I will release a tiny vlm( t1-vl-grpo ).