diff --git a/README.md b/README.md
index 8f2ab29..8328a1c 100644
--- a/README.md
+++ b/README.md
@@ -18,21 +18,31 @@ pip3 install flash-attn --no-build-isolation
 pip install wandb IPython matplotlib
 ```
 
+## The Task
+
 ## Generate Data
 ```
 conda activate zero
-python verl/examples/data_preprocess/countdown.py
+python examples/data_preprocess/countdown.py
 ```
 
 ## Run Training
 ```
 conda activate zero
 ```
+**Single GPU dry run**
+Works for model <= 1.5B
+
+For Qwen2.5-0.5B base, we know it fails to learn reasoning.
+
 ```
 export CUDA_VISIBLE_DEVICES=7
 export N_GPUS=1
-export BASE_MODEL=Qwen/Qwen2.5-0.5B
+export BASE_MODEL=Qwen/Qwen2.5-1.5B
 export DATA_DIR=$HOME/data/countdown
+export WANDB_API_KEY=0929e692448f1bc929d71d7e3ece80073c3041e6
+export EXPERIMENT_NAME=countdown-qwen2.5-1.5b
+
 PYTHONUNBUFFERE=1 python3 -m verl.trainer.main_ppo \
  data.train_files=$DATA_DIR/train.parquet \
  data.val_files=$DATA_DIR/test.parquet \
@@ -57,9 +67,93 @@ PYTHONUNBUFFERE=1 python3 -m verl.trainer.main_ppo \
  trainer.default_hdfs_dir=null \
  trainer.n_gpus_per_node=$N_GPUS \
  trainer.nnodes=1 \
- trainer.save_freq=10 \
+ trainer.save_freq=30 \
  trainer.test_freq=10 \
- trainer.project_name=zero \
- trainer.experiment_name=countdown \
+ trainer.project_name=TinyZero \
+ trainer.experiment_name=$EXPERIMENT_NAME \
  trainer.total_epochs=15 2>&1 | tee verl_demo.log
 ```
+
+**3B model dry run**
+In this case, the base model is able to develop sophisticated reasoning skills.
+```
+export CUDA_VISIBLE_DEVICES=4,5
+export N_GPUS=2
+export BASE_MODEL=Qwen/Qwen2.5-3B
+export DATA_DIR=$HOME/data/countdown
+export ROLLOUT_TP_SIZE=2
+export WANDB_API_KEY=0929e692448f1bc929d71d7e3ece80073c3041e6
+export EXPERIMENT_NAME=countdown-qwen2.5-3b
+
+python3 -m verl.trainer.main_ppo \
+ data.train_files=$DATA_DIR/train.parquet \
+ data.val_files=$DATA_DIR/test.parquet \
+ data.train_batch_size=256 \
+ data.val_batch_size=1312 \
+ data.max_prompt_length=256 \
+ data.max_response_length=1024 \
+ actor_rollout_ref.model.path=$BASE_MODEL \
+ actor_rollout_ref.actor.optim.lr=1e-6 \
+ actor_rollout_ref.actor.ppo_mini_batch_size=128 \
+ actor_rollout_ref.actor.ppo_micro_batch_size=8 \
+ actor_rollout_ref.rollout.log_prob_micro_batch_size=8 \
+ actor_rollout_ref.rollout.tensor_model_parallel_size=$ROLLOUT_TP_SIZE \
+ actor_rollout_ref.rollout.gpu_memory_utilization=0.4 \
+ actor_rollout_ref.ref.log_prob_micro_batch_size=4 \
+ critic.optim.lr=1e-5 \
+ critic.model.path=$BASE_MODEL \
+ critic.ppo_micro_batch_size=8 \
+ algorithm.kl_ctrl.kl_coef=0.001 \
+ trainer.logger=['wandb'] \
+ +trainer.val_before_train=False \
+ trainer.default_hdfs_dir=null \
+ trainer.n_gpus_per_node=$N_GPUS \
+ trainer.nnodes=1 \
+ trainer.save_freq=30 \
+ trainer.test_freq=10 \
+ trainer.project_name=TinyZero \
+ trainer.experiment_name=$EXPERIMENT_NAME \
+ trainer.total_epochs=15 2>&1 | tee verl_demo.log
+```
+
+**OpenLlama 7B model dry run**
+In this case, the base model is able to develop sophisticated reasoning skills.
+```
+export CUDA_VISIBLE_DEVICES=4,5,6,7
+export N_GPUS=4
+export EXPERIMENT_NAME=countdown-open_llama_7b
+export BASE_MODEL=openlm-research/open_llama_7b_v2
+export DATA_DIR=$HOME/data/countdown
+export ROLLOUT_TP_SIZE=4
+export WANDB_API_KEY=0929e692448f1bc929d71d7e3ece80073c3041e6
+
+python3 -m verl.trainer.main_ppo \
+ data.train_files=$DATA_DIR/train.parquet \
+ data.val_files=$DATA_DIR/test.parquet \
+ data.train_batch_size=256 \
+ data.val_batch_size=1312 \
+ data.max_prompt_length=256 \
+ data.max_response_length=1024 \
+ actor_rollout_ref.model.path=$BASE_MODEL \
+ actor_rollout_ref.actor.optim.lr=1e-6 \
+ actor_rollout_ref.actor.ppo_mini_batch_size=128 \
+ actor_rollout_ref.actor.ppo_micro_batch_size=8 \
+ actor_rollout_ref.rollout.log_prob_micro_batch_size=8 \
+ actor_rollout_ref.rollout.tensor_model_parallel_size=$ROLLOUT_TP_SIZE \
+ actor_rollout_ref.rollout.gpu_memory_utilization=0.4 \
+ actor_rollout_ref.ref.log_prob_micro_batch_size=4 \
+ critic.optim.lr=1e-5 \
+ critic.model.path=$BASE_MODEL \
+ critic.ppo_micro_batch_size=8 \
+ algorithm.kl_ctrl.kl_coef=0.001 \
+ trainer.logger=['wandb'] \
+ +trainer.val_before_train=False \
+ trainer.default_hdfs_dir=null \
+ trainer.n_gpus_per_node=$N_GPUS \
+ trainer.nnodes=1 \
+ trainer.save_freq=30 \
+ trainer.test_freq=10 \
+ trainer.project_name=TinyZero \
+ trainer.experiment_name=$EXPERIMENT_NAME \
+ trainer.total_epochs=15 2>&1 | tee verl_demo.log
+```
\ No newline at end of file
diff --git a/verl/utils/tracking.py b/verl/utils/tracking.py
index 809a8a2..b1fbd6f 100644
--- a/verl/utils/tracking.py
+++ b/verl/utils/tracking.py
@@ -38,6 +38,10 @@ def __init__(self, project_name, experiment_name, default_backend: Union[str, Li
 
         if 'tracking' in default_backend or 'wandb' in default_backend:
             import wandb
+            import os
+            WANDB_API_KEY = os.environ.get("WANDB_API_KEY", None)
+            if WANDB_API_KEY:
+                wandb.login(key=WANDB_API_KEY)
             wandb.init(project=project_name, name=experiment_name, config=config)
             self.logger['wandb'] = wandb