docs: add Docker installation guide for OpenJudge and training enviro…

docker/Dockerfile.base

@@ -0,0 +1,49 @@
+# OpenJudge Dockerfile
+# Base image for running OpenJudge evaluation tasks
+# For training Judge models with verl/sglang/vllm, use Dockerfile.train instead
+
+FROM dsw-registry.cn-wulanchabu.cr.aliyuncs.com/pai/pytorch:2.8.0-gpu-py312-cu128-ubuntu24.04-3995b779-1764359181
+
+# Set environment variables
+ENV PYTHONUNBUFFERED=1
+ENV DEBIAN_FRONTEND=noninteractive
+
+# Set working directory
+WORKDIR /workspace
+
+# Install system dependencies
+RUN apt-get update && apt-get install -y --no-install-recommends \
+    git \
+    curl \
+    build-essential \
+    && rm -rf /var/lib/apt/lists/*
+
+# Install OpenJudge with all dependencies
+RUN pip install --no-cache-dir \
+    "pandas>=2.2.3,<3.0.0" \
+    "loguru>=0.7.3,<0.8.0" \
+    "json_repair>=0.54.0,<1.0.0" \
+    "pydantic>=2.11.5,<3.0.0" \
+    "openai>=1.85.0,<2.0.0" \
+    "tenacity>=9.1.0,<10.0.0" \
+    "math-verify>=0.7.0,<0.8.0" \
+    "tqdm>=4.66.0,<5.0.0" \
+    "fire" \
+    "numpy>=1.22.0,<2.0.0" \
+    "dashscope>=1.19.0" \
+    "tiktoken>=0.7.0" \
+    "nltk>=3.8.1" \
+    "jieba>=0.42.1" \
+    "sacrebleu>=2.0.0" \
+    "rouge-score>=0.1.2" \
+    "python-Levenshtein>=0.20.0" \
+    "scikit-learn>=1.0.0"
+
+# Install OpenJudge from GitHub
+RUN pip install --no-cache-dir git+https://github.com/agentscope-ai/OpenJudge.git
+
+# Clean up
+RUN pip cache purge
+
+# Set default command
+CMD ["/bin/bash"]

docker/Dockerfile.train

@@ -0,0 +1,123 @@
+# =============================================================================
+# OpenJudge Training Dockerfile
+# =============================================================================
+# Training image for Judge model SFT/RL with:
+#   - PyTorch + CUDA support
+#   - Inference frameworks: sglang, vllm
+#   - Training frameworks: verl, transformers, accelerate
+#   - FlashAttention, FlashInfer
+#   - OpenJudge
+#
+# For basic installation (evaluation only), use Dockerfile instead
+# =============================================================================
+
+# Base image: PAI PyTorch
+FROM dsw-registry.cn-wulanchabu.cr.aliyuncs.com/pai/pytorch:2.8.0-gpu-py312-cu128-ubuntu24.04-3995b779-1764359181
+
+# Set environment variables
+ENV USE_MEGATRON=0
+ENV USE_SGLANG=1
+ENV MAX_JOBS=32
+ENV DEBIAN_FRONTEND=noninteractive
+
+# Set working directory
+WORKDIR /workspace
+
+# =============================================================================
+# 1. Inference Frameworks (sglang, vllm)
+# =============================================================================
+RUN pip install --no-cache-dir "sglang[all]==0.5.2" && \
+    pip install --no-cache-dir torch-memory-saver && \
+    pip install --no-cache-dir "vllm==0.11.0"
+
+# =============================================================================
+# 2. Training & ML Packages
+# =============================================================================
+RUN pip install --no-cache-dir \
+    "transformers[hf_xet]>=4.51.0" \
+    accelerate \
+    datasets \
+    peft \
+    hf-transfer \
+    "numpy<2.0.0" \
+    "pyarrow>=15.0.0" \
+    pandas \
+    "tensordict>=0.8.0,<=0.10.0,!=0.9.0" \
+    torchdata \
+    "ray[default]" \
+    codetiming \
+    hydra-core \
+    pylatexenc \
+    qwen-vl-utils \
+    wandb \
+    swanlab \
+    dill \
+    pybind11 \
+    liger-kernel \
+    mathruler \
+    pytest \
+    py-spy \
+    pre-commit \
+    ruff \
+    tensorboard
+
+# =============================================================================
+# 3. Additional Dependencies
+# =============================================================================
+RUN pip install --no-cache-dir \
+    "nvidia-ml-py>=12.560.30" \
+    "fastapi[standard]>=0.115.0" \
+    "optree>=0.13.0" \
+    "pydantic>=2.9" \
+    "grpcio>=1.62.1"
+
+# =============================================================================
+# 4. FlashAttention & FlashInfer (Python 3.12 + CUDA 12)
+# =============================================================================
+RUN curl -L -O "https://github.com/Dao-AILab/flash-attention/releases/download/v2.8.1/flash_attn-2.8.1+cu12torch2.8cxx11abiFALSE-cp312-cp312-linux_x86_64.whl" && \
+    pip install --no-cache-dir flash_attn-2.8.1+cu12torch2.8cxx11abiFALSE-cp312-cp312-linux_x86_64.whl && \
+    rm -f flash_attn-2.8.1+cu12torch2.8cxx11abiFALSE-cp312-cp312-linux_x86_64.whl
+
+RUN pip install --no-cache-dir flashinfer-python==0.3.1
+
+# =============================================================================
+# 5. OpenCV Fix
+# =============================================================================
+RUN pip install --no-cache-dir opencv-python opencv-fixer && \
+    python -c "from opencv_fixer import AutoFix; AutoFix()"
+
+# =============================================================================
+# 6. verl (RL Training Framework)
+# =============================================================================
+RUN pip install --no-cache-dir --no-deps git+https://github.com/volcengine/[email protected]
+
+# =============================================================================
+# 7. OpenJudge Dependencies
+# =============================================================================
+RUN pip install --no-cache-dir \
+    "loguru>=0.7.3,<0.8.0" \
+    "json_repair>=0.54.0,<1.0.0" \
+    "openai>=1.85.0,<2.0.0" \
+    "tenacity>=9.1.0,<10.0.0" \
+    "math-verify>=0.7.0,<0.8.0" \
+    "tqdm>=4.66.0,<5.0.0" \
+    fire \
+    "dashscope>=1.19.0" \
+    "tiktoken>=0.7.0" \
+    "nltk>=3.8.1" \
+    "jieba>=0.42.1" \
+    "sacrebleu>=2.0.0" \
+    "rouge-score>=0.1.2" \
+    "python-Levenshtein>=0.20.0" \
+    "scikit-learn>=1.0.0"
+
+# =============================================================================
+# 8. OpenJudge
+# =============================================================================
+RUN pip install --no-cache-dir --no-deps git+https://github.com/agentscope-ai/OpenJudge.git
+
+# Clean up cache
+RUN pip cache purge
+
+# Set default command
+CMD ["/bin/bash"]

docker/README.md

@@ -0,0 +1,84 @@
+# OpenJudge Docker Guide
+
+This document describes how to deploy OpenJudge using Docker. We provide two images:
+
+| Image | Purpose | Dockerfile |
+|-------|---------|------------|
+| **Base Image** | Running evaluation tasks (API calls) | `Dockerfile.base` |
+| **Training Image** | Judge model SFT/RL training | `Dockerfile.train` |
+
+---
+
+## Option 1: Base Installation (Evaluation)
+
+For scenarios using OpenJudge for evaluation (calling LLMs via API).
+
+### 1.1 Build Image
+
+```bash
+cd OpenJudge
+docker build -f docker/Dockerfile.base -t openjudge:latest .
+```
+
+### 1.2 Start Container
+
+```bash
+docker run -it \
+    -v $(pwd):/workspace/OpenJudge \
+    -e OPENAI_API_KEY=your_api_key \
+    --name openjudge \
+    openjudge:latest
+```
+
+---
+
+## Option 2: Training Environment
+
+For scenarios using the [verl](https://github.com/volcengine/verl) framework for Judge model SFT/RL training.
+
+### Environment Details
+
+- **Base Image**: PAI PyTorch 2.8.0 + CUDA 12.8 + Python 3.12
+- **Training Framework**: verl v0.6.1 (FSDP distributed training)
+- **Inference Frameworks**: vLLM 0.11.0, SGLang 0.5.2
+
+### 2.1 Build Image
+
+```bash
+cd OpenJudge
+docker build -f docker/Dockerfile.train -t openjudge-train:latest .
+```
+
+### 2.2 Start Container
+
+```bash
+docker run --gpus all -it \
+    --shm-size=64g \
+    -v $(pwd):/workspace/OpenJudge \
+    -v /path/to/your/models:/models \
+    -v /path/to/your/data:/data \
+    --name openjudge-train \
+    openjudge-train:latest
+```
+
+**Parameter Description:**
+
+| Parameter | Description |
+|-----------|-------------|
+| `--gpus all` | Use all GPUs |
+| `--shm-size=64g` | Set shared memory to 64GB (required for training) |
+| `-v $(pwd):/workspace/OpenJudge` | Mount current directory to container |
+| `-v /path/to/your/models:/models` | Mount model directory (modify path as needed) |
+| `-v /path/to/your/data:/data` | Mount data directory (modify path as needed) |
+| `--name` | Container name |
+
+### 2.3 Run Training
+
+After entering the container:
+
+```bash
+cd /workspace/OpenJudge/cookbooks/training_judge_model/sft
+bash run_sft_rm.sh
+```
+
+---

docs/building_graders/training_judge_models.md

@@ -16,11 +16,51 @@ OpenJudge provides training pipelines for building custom judge models. Each met
 | **Bradley-Terry** | Scalar score | Preference pairs | ❌ No | RLHF judge modeling, ranking |
 | **GRPO** | Generative (text) | Labeled responses | ✅ Yes | Interpretable evaluation with reasoning |
 
-**Common Requirements:**
+**Environment Setup:**
 
-```bash
-pip install verl==0.6.1
-```
+=== "Docker (Recommended)"
+
+    Use the pre-configured training Docker image:
+
+    ```bash
+    cd OpenJudge
+
+    # Build the training image
+    docker build -f docker/Dockerfile.train -t openjudge-train:latest .
+
+    # Run the container
+    docker run --gpus all -it \
+        --shm-size=64g \
+        -v $(pwd):/workspace/OpenJudge \
+        -v /path/to/your/models:/models \
+        -v /path/to/your/data:/data \
+        --name openjudge-train \
+        openjudge-train:latest
+    ```
+
+=== "Manual Installation"
+
+    For custom environments, follow the [verl installation guide](https://verl.readthedocs.io/en/latest/start/install.html):
+
+    ```bash
+    # 1. Create conda environment
+    conda create -n openjudge_train python==3.12
+    conda activate openjudge_train
+
+    # 2. Install dependencies (vLLM, SGLang, etc.)
+    git clone https://github.com/volcengine/verl && cd verl
+    USE_MEGATRON=0 bash scripts/install_vllm_sglang_mcore.sh
+
+    # 3. Install verl
+    pip install --no-deps -e .
+
+    # 4. Install OpenJudge (verl dependencies already installed in step 2)
+    cd /path/to/OpenJudge
+    pip install -e .
+    ```
+
+    !!! warning "Important"
+        Simply running `pip install verl` is **not sufficient**. verl requires specific versions of vLLM, SGLang, FlashAttention, and other dependencies. Please follow the full installation guide or use the Docker image.
 
 
 ## Datasets

docs/get_started/quickstart.md

@@ -26,6 +26,22 @@ Get started with OpenJudge in 5 minutes. This guide walks you through installati
     pip install -e .[verl]  # With VerL option for training scenarios
     ```
 
+=== "Docker"
+
+    ```bash
+    git clone https://github.com/agentscope-ai/OpenJudge.git
+    cd OpenJudge
+
+    # Build the Docker image
+    docker build -f docker/Dockerfile.base -t openjudge:latest .
+
+    # Run the container
+    docker run -it \
+        -v $(pwd):/workspace/OpenJudge \
+        --name openjudge \
+        openjudge:latest
+    ```
+
 > **Tips:**
 > OpenJudge requires Python version >=3.10 and <3.13. For best compatibility, we recommend using Python 3.10 or 3.11.