diff --git a/examples/multi-agent/run_train_chain.sh b/examples/multi-agent/run_train_chain.sh
old mode 100644
new mode 100755
index 1e59da3..6c698e0
--- a/examples/multi-agent/run_train_chain.sh
+++ b/examples/multi-agent/run_train_chain.sh
@@ -6,17 +6,28 @@
 # cd /mnt/shared-storage-user/marti/OpenRLHF
 
 # 基础配置
-MODEL_DIR="/your_model_path"
+MODEL_DIR="Qwen"
 SHORT_NAME=${1:-"Qwen2.5-3B-Instruct"}
 PRETRAIN="${MODEL_DIR}/${SHORT_NAME}"
 PROMPT_MAX_LEN=4096
 GENERATE_MAX_LEN=32768
 EVAL_GENERATE_MAX_LEN=32768
 OVERLONG_BUFFER_LEN=2048
-MAX_LEN=42024 # 30000
+MAX_LEN=30000 #42024
 ADVANTAGE="group_norm"
 
-ROOT_DIR=""
+# Repo root: use ROOT_DIR if set, else derive from this script's path
+if [ -z "${ROOT_DIR}" ]; then
+  SCRIPT_DIR="$(cd "$(dirname "${BASH_SOURCE[0]:-$0}")" && pwd)"
+  ROOT_DIR="$(cd "${SCRIPT_DIR}/../.." && pwd)"
+fi
+[ -n "${ROOT_DIR}" ] || ROOT_DIR="$(pwd)"
+
+# Hugging Face cache under repo (avoids PermissionError on ~/.cache/huggingface)
+export HF_DATASETS_CACHE="${ROOT_DIR}/.cache/huggingface/datasets"
+export HF_HOME="${ROOT_DIR}/.cache/huggingface"
+mkdir -p "$HF_DATASETS_CACHE"
+
 TASK="MATH"
 PROMPT_DATA="json@${ROOT_DIR}/data/${TASK}"
 
@@ -104,12 +115,12 @@ mkdir -p "${ROOT_DIR}/outputs/ckpt/${ADVANTAGE}-${SHORT_NAME}-${TASK}-ch-${EXP}-
     # --reward_alloc "${REWARD_ALLOC_ARGS}" \
 
 
-python3 -m openrlhf.cli.multi_agent_train_ppo_ray \
+python3 -m marti.cli.multi_agent_train_ppo_ray \
     --default_agent "$DEFAULT_AGENT" \
     --agents "$AGENT0" "$AGENT1" "$AGENT2" \
     --workflow_args "$WORKFLOW_ARGS" \
-    --workflow_func_path openrlhf/agent_workflows/chain_workflow.py \
-    --processor_func_path openrlhf/agent_workflows/chain_processor.py \
+    --workflow_func_path marti/agent_workflows/chain_workflow.py \
+    --processor_func_path marti/agent_workflows/chain_processor.py \
     --parallel_loading \
     --ref_num_nodes 1 \
     --ref_num_gpus_per_node 1 \
diff --git a/marti/trainer/ray/vllm_engine.py b/marti/trainer/ray/vllm_engine.py
index e4cdbb5..0362873 100755
--- a/marti/trainer/ray/vllm_engine.py
+++ b/marti/trainer/ray/vllm_engine.py
@@ -172,7 +172,7 @@ def create_vllm_engines(
             ).remote(
                 model=pretrain,
                 enforce_eager=enforce_eager,
-                worker_extension_cls="openrlhf.trainer.ray.vllm_worker_wrap.WorkerWrap",
+                worker_extension_cls="marti.trainer.ray.vllm_worker_wrap.WorkerWrap",
                 tensor_parallel_size=tensor_parallel_size,
                 seed=seed + i,
                 distributed_executor_backend=distributed_executor_backend,
diff --git a/requirements.txt b/requirements.txt
index b41e72b..fb6acb5 100755
--- a/requirements.txt
+++ b/requirements.txt
@@ -6,20 +6,26 @@ einops
 flash-attn==2.8.3
 grpcio>=1.74.0
 isort
+json5
 jsonlines
+latex2sympy2
 loralib
 optimum
 optree>=0.13.0
 packaging
 peft
+pylatexenc
 pynvml>=12.0.0
 ray[default]==2.48.0
+srsly
 tensorboard
-torch
+torch==2.9
 torchdata
 torchmetrics
 tqdm
 transformers==4.57.0
 transformers_stream_generator
+vllm>0.8.5.post1
 wandb
 wheel
+word2number