Ostrakon-VL-8B

Ostrakon-VL: Towards Domain-Expert MLLM for Food-Service and Retail Stores]

Introduction

Ostrakon-VL is the first open-source Multimodal Large Language Model (MLLM) specifically designed for Food-Service and Retail Store (FSRS) scenarios. It delivers state-of-the-art performance in real-world retail perception, compliance, and decision-making tasks — outperforming even much larger general-purpose models like Qwen3-VL-235B.

🔥 Key Highlights

• 60.5 average score on ShopBench (Ostrakon-VL-30B-A3B) — outperforming Qwen3-VL-235B by +1.1
• ShopBench: First public benchmark for FSRS, with multi-format inputs and diagnostic metrics
• Model weights are open-sourced

---

📦 What’s Included

• ✅ Ostrakon-VL Models: Fine-tuned MLLMs for FSRS, including 8B and 30B variants (ModelScope: https://modelscope.cn/collections/Ostrakon/Ostrakon-VL)
• 🧪 ShopBench: The first public benchmark for FSRS, featuring:
  • Scenarios: ShopFront, ShopInterior, Kitchen
  • Input Types: Single image, multi-image, video
  • Output Formats: Open-ended QA, structured Format, MCQ
  • Unique Design: High visual complexity (13.0 objects/image), fine-grained task taxonomy (79 categories), and diagnostic metrics (VNR/VIF) to minimize language bias

---

🧪 ShopBench: The First Public Benchmark for FSRS

ShopBench is the first standardized evaluation suite specifically designed for the Food-Service and Retail Store (FSRS) domain. Unlike general-purpose multimodal benchmarks, ShopBench addresses the unique challenges of real-world retail environments:

• Multi-format Inputs: Unifies evaluation across single-image, multi-image, and video scenarios under a consistent protocol.
• Fine-grained Taxonomy: Organizes tasks into a 4-level hierarchy covering perception (e.g., OCR, object localization) and reasoning (e.g., spatial relations, compliance rules), with 79 leaf-level categories.
• Real-world Complexity: Features the highest instance density (13.0 objects/image) among major benchmarks, reflecting cluttered retail scenes.
• Reduced Language Bias: Employs diagnostic metrics like Visual Necessity Rate (VNR) and Vision-Induced Failure (VIF) to ensure questions truly require visual understanding, not just language priors.
• Domain Distinctiveness: t-SNE visualization shows ShopBench occupies a distinct region in the visual embedding space compared to existing benchmarks (e.g., MMBench, OCRBench).

ShopBench enables fair, interpretable, and actionable comparison of MLLMs in realistic FSRS workflows — from storefront compliance checks to kitchen hygiene audits.

Due to legal reasons, the open-sourcing of ShopBench has been postponed.

---

📊 Evaluation Results

ShopBench Performance (Higher is better)

Our Ostrakon-VL models achieve strong performance on ShopBench. Ostrakon-VL-8B reaches an average score of 60.1, showing competitive results among open-source models of similar scale, while Ostrakon-VL-30B-A3B further improves to 60.5 and performs particularly well on ShopFront (65.3) and Video (57.3), demonstrating strong understanding of e-commerce scenarios.

General Benchmarks

While specialized for FSRS, On representative multimodal benchmarks, Ostrakon-VL shows competitive overall performance across a wide range of tasks. In particular, Ostrakon-VL-30B-A3B achieves strong results on MMBench-EN_DEV (86.7) and AI2D_TEST (86.9), while maintaining solid performance across OCR, math, and Chinese-language benchmarks, demonstrating its broad multimodal capabilities.

---

🧠 Training Strategy

Ostrakon-VL is trained via a multi-stage training strategy:

1. Caption Bootstrapping (CB) – Inject domain knowledge using dense, evidence-rich captions.
2. Offline Curriculum Learning (OCL) – Train from easy to hard samples based on difficulty stratification.
3. Mixed Preference Optimization (MPO) – Align responses using preference pairs of correct vs. plausible-but-wrong answers.

This strategy enables robust, auditable, and rule-compliant reasoning in complex retail environments.

---

🚀 Quick Start

Install dependencies and run inference with just a few lines of code:

pip install torch transformers accelerate

Inference Example

from transformers import Qwen3VLForConditionalGeneration, AutoProcessor
from PIL import Image

# Load Ostrakon-VL
model = Qwen3VLForConditionalGeneration.from_pretrained(
    "Ostrakon/Ostrakon-VL-30B-A3B",
    dtype="bfloat16",
    device_map="auto",
)
processor = AutoProcessor.from_pretrained("Ostrakon/Ostrakon-VL-30B-A3B")

def process_local_image(image_path):
    image = Image.open(image_path)
    image = image.resize((512,512))
    return image

image_path = "path/to/img.jpg"
image = process_local_image(image_path)

messages = [
    {
        "role": "user",
        "content": [
            {
                "type": "image",
                "image": image
            },
            {
                "type": "text",
                "text": "图片中的店铺名是什么？"
            }
        ]
    }
]

inputs = processor.apply_chat_template(
    messages,
    tokenize=True,
    add_generation_prompt=True,
    return_dict=True,
    return_tensors="pt"
)
inputs = inputs.to(model.device)

generated_ids = model.generate(**inputs, max_new_tokens=128)
generated_ids_trimmed = [
    out_ids[len(in_ids) :] for in_ids, out_ids in zip(inputs.input_ids, generated_ids)
]
output_text = processor.batch_decode(
    generated_ids_trimmed, skip_special_tokens=True, clean_up_tokenization_spaces=False
)[0]
print(output_text)

📚 Citation

If you find Ostrakon-VL or ShopBench useful in your research, please cite our paper:

@article{shen2026ostrakon,
  title={Ostrakon-VL: Towards Domain-Expert MLLM for Food-Service and Retail Stores},
  author={Shen, Zhiyong and Zhao, Gongpeng and Zhou, Jun and Yu, Li and Kou, Guandong and Li, Jichen and Dong, Chuanlei and Li, Zuncheng and Li, Kaimao and Wei, Bingkun and others},
  journal={arXiv preprint arXiv:2601.21342},
  year={2026}
}

---

🤝 Contributing & License

We welcome contributions! Please open an issue or submit a PR.

• Code & Model: Apache 2.0 License