unsatisfactory result and strange reward

[qianfantianyuzhouzhou]

it seems that the table's reward only denpends on format reward?

and my training result( MATH dataset on Qwen2.5-3B with GRPO, only 37%) is not good enough?

here is my parameters: the same as example:

[qianfantianyuzhouzhou]

bf16: true
use_vllm: true
vllm_device: auto
vllm_gpu_memory_utilization: 0.6
do_eval: true
eval_strategy: steps
eval_steps: 100
gradient_accumulation_steps: 8
gradient_checkpointing: true
gradient_checkpointing_kwargs:
use_reentrant: false
hub_model_id: Qwen-2.5-3B
learning_rate: 3.0e-06
lr_scheduler_type: cosine
max_prompt_length: 512
max_completion_length: 1024
max_steps: -1
num_generations: 6
num_train_epochs: 1
output_dir: data/Qwen-2.5-3B-Simple-RL
overwrite_output_dir: true
per_device_eval_batch_size: 2
per_device_train_batch_size: 2
push_to_hub: false
report_to:

• wandb
  reward_funcs:
• accuracy
• format
  reward_weights:
• 1.0
• 1.0
  save_strategy: "no"
  seed: 42
  warmup_ratio: 0.1