Agnes-SeaLLM-8B

Introduction

We introduce Agnes-SeaLLM-8B, a compact Large Language Model (LLM) meticulously optimized for Southeast Asian languages. Despite its efficient footprint, it delivers performance rivaling much larger open-source models, excelling in tasks such as mathematical reasoning, translation, and instruction following. Furthermore, it has been specially tuned to enhance reliability, minimize hallucinations, and provide culturally sensitive, safe responses tailored to the Southeast Asian context.

🔥 Highlights

Compact Efficiency & Rapid Deployment: Significantly smaller than mainstream LLMs, Agnes-SeaLLM-8B enables high-speed inference and low-resource deployment without sacrificing accuracy or multilingual proficiency. It is the ideal choice for edge devices and resource-constrained environments.

Top-Tier Performance in its Class: Outperforms comparable open-source models across multi-dimensional benchmarks, including academic examinations, complex instruction following, mathematics, and high-precision translation.

Superior Instruction Following: Demonstrates exceptional capability in handling multi-turn dialogues and executing complex, nuanced tasks with high fidelity.

Safe, Reliable, and Culturally Aware: Engineered to reduce hallucinations and increase sensitivity to Southeast Asian cultural nuances. This ensures safer interactions while maintaining peak performance in English and Chinese.

Balanced Multilingual Mastery: Maintains high-quality, consistent output across a broad linguistic spectrum, overcoming the "seesaw effect" often found in regional models.

【SeaExam Summary statistics】

• Agnes-SeaLLM-8B (75.32 avg) crushes 8B peers and outpaces Sailor2-20B, offering 70B-tier performance with compact, low-resource deployment.
• Exceptional scores in both M3Exam and MMLU prove it bridges regional linguistic nuance with elite global reasoning.

【M3Exam Model evaluation sub-statistics】

• With a 71.29 average, Agnes-SeaLLM-8B delivers "70B-class" performance in an 8B footprint, outperforming the 20B Sailor2 and crushing 8B peers like Llama-3 and GLM-4.
• It sets a new benchmark for Southeast Asia, achieving top-tier scores in Indonesian (72.97) and Vietnamese (70.03)**, while maintaining elite English (93.24) and Chinese (83.70) proficiency.

【MMLU Model evaluation sub-statistics】

【Indonesian Fine-grained discipline statistics】

【MMLU Fine-grained discipline statistics】

📦 SeaExam Evaluation Package

/
├── function.py      # Main function 
├── main.py           # Main evaluation script
├── requirements.txt      # Python dependencies
├── data/                 # Test datasets (26MB)
│   ├── m3exam-chinese/
│   ├── m3exam-english/
│   ├── m3exam-indonesian/
│   ├── m3exam-thai/
│   ├── m3exam-vietnamese/
│   ├── mmlu-chinese/
│   ├── mmlu-english/
│   ├── mmlu-indonesian/
│   ├── mmlu-thai/
│   └── mmlu-vietnamese/
├── models/               # [YOU NEED TO ADD] Place your model here
│   └── Llama-SEA-LION-v2-8B-IT/  (or your model name)
├── README.md             # This file
└── quick_run.sh                # Quick start script

🔧 Prerequisites

Hardware Requirements

• GPU: Minimum 2 GPUs with 24GB+ VRAM each
• RAM: 64GB+ recommended
• Storage: ~20GB (model + data + outputs)

Software Requirements

• OS: Linux (tested on Ubuntu)
• Python: 3.10 - 3.12
• CUDA: 12.x
• GPU Driver: Latest NVIDIA driver

📥 Installation Steps

1. Copy this folder to your server

# Example: Using scp
scp -r Seaexam/ user@your-server:/path/to/destination/

# Or using rsync
rsync -avz Seaexam/ user@your-server:/path/to/destination/Seaexam/

2. Prepare your model

Place your model files in the models/ directory:

cd Seaexam
mkdir -p models
# Copy or download your model to models/Llama-SEA-LION-v2-8B-IT/
# The model directory should contain:
# - config.json
# - tokenizer files
# - model weight files (.safetensors or .bin)

3. Create Python virtual environment (recommended)

cd Seaexam
python3 -m venv venv
source venv/bin/activate  # On Windows: venv\Scripts\activate

4. Install dependencies

conda create -n SeaExam python=3.12
conda activate SeaExam
pip install --upgrade pip
pip install -r requirements.txt

Note: This will install PyTorch, vLLM, and other dependencies. May take 10-20 minutes.

⚙️ Configuration

Method 1: Edit main.py directly

Open main.py and modify the configuration section (lines 12-24):

# Set your model path
MODEL_NAME = "./models/Llama-SEA-LION-v2-8B-IT"  # Or use absolute path

# GPU settings
TENSOR_PARALLEL_SIZE = 2  # Number of GPUs (must match CUDA_VISIBLE_DEVICES)
os.environ["CUDA_VISIBLE_DEVICES"] = "0,1"  # GPU IDs to use

# Batch size (adjust based on your GPU memory)
BATCH_SIZE = 256

Method 2: Use environment variable

export MODEL_PATH="/path/to/your/model"
export CUDA_VISIBLE_DEVICES="0,1"
python main.py

Configuration Tips

GPU Selection

• Check available GPUs: nvidia-smi
• Use specific GPUs: CUDA_VISIBLE_DEVICES="0,1" (uses GPU 0 and 1)
• Match TENSOR_PARALLEL_SIZE with number of GPUs in CUDA_VISIBLE_DEVICES

Memory Optimization

If you encounter OOM (Out of Memory) errors:

1. Reduce BATCH_SIZE (e.g., 128, 64, 32)
2. Reduce gpu_memory_utilization in code (default 0.95 → 0.8)
3. Reduce max_model_len (default 4096 → 2048)

🚀 Running the Evaluation

Quick Start

cd Seaexam
source venv/bin/activate  # If using virtual environment
python main.py

Using the run script

chmod +x quick_run.sh
./quick_run.sh

Monitor Progress

The script shows a progress bar with real-time accuracy:

Inference: 45% |████████          | 2345/5000 [10:23<12:45] Acc: 67.23%

Resume from Checkpoint

The script automatically saves progress to progress.json. If interrupted:

• Just run the script again
• It will resume from where it stopped

📊 Output Files

After completion, you will find:

• result.txt - Detailed evaluation report with accuracy by language/subject
• progress.json - Progress checkpoint (can be deleted after successful completion)

🔍 Verification

Before running the full evaluation, verify your setup:

# Check CUDA installation
nvcc --version

# Check GPU availability
nvidia-smi

# Check Python packages
pip list | grep -E "torch|transformers|vllm"

# Test model loading (Python)
python -c "from transformers import AutoTokenizer; tokenizer = AutoTokenizer.from_pretrained('./models/Llama-SEA-LION-v2-8B-IT', trust_remote_code=True); print('Model loads successfully!')"

📝 Expected Results

• Total Questions: ~5000 (varies by data version)
• Runtime: 1-3 hours (depends on hardware)
• Output Accuracy: Model-dependent (typically 40-80%)

❓ Troubleshooting

Issue: "No module named 'vllm'"

Solution: Install dependencies: pip install -r requirements.txt

Issue: "CUDA out of memory"

Solution:

1. Use fewer GPUs but ensure VRAM > 24GB per GPU
2. Check GPU usage: nvidia-smi

Issue: "No test data files found"

Solution: Verify data/ directory structure and ensure test.json files exist

Issue: CUDA version mismatch

Solution:

# Check CUDA version
nvcc --version

# If CUDA < 12.x, update requirements.txt:
# Replace cupy-cuda12x with cupy-cuda11x
# Replace nvidia-cuda-* packages with appropriate versions

📚 Additional Information

Model Format Support

• Hugging Face format (recommended)
• SafeTensors format
• PyTorch .bin format

Data Format

Each test.json contains:

{
  "question": "...",
  "choices": ["A", "B", "C", "D"],
  "answer": 0,  // Index of correct answer (0-3)
  "metadata": {
    "subject": "...",
    "language": "..."
  }
}

Customization

You can modify the evaluation prompt in the build_prompt() function.

📧 Support

If you encounter issues:

1. Check the troubleshooting section above
2. Verify all prerequisites are met
3. Check GPU/CUDA compatibility
4. Review error messages carefully

🔄 Version Information

• Package Version: 1.0
• Python: 3.10-3.12
• PyTorch: 2.8.0
• vLLM: 0.11.0
• Transformers: 4.57.1
• CUDA: 12.x