README.md

---
license: cc-by-4.0
pipeline_tag: image-text-to-text
language:
  - tr
  - en
base_model:
  - Qwen/Qwen3.5-27B
model_type: vision-language-model
tags:
  - computer-vision
  - multimodal
  - e-commerce
  - catalog-moderation
  - vision-language-model
  - product-understanding
library_name: transformers
---

# Trendyol-Vision-Master

_Trendyol-Vision-Master is a fine-tuned **vision-language model (VLM)** built on [Qwen/Qwen3.5-27B](https://huggingface.co/Qwen/Qwen3.5-27B) for Trendyol catalog quality and moderation workflows. It understands product images, titles, and metadata, and produces structured decisions or text outputs for **e-commerce catalog operations** — including category detection, brand verification, product similarity, attribute extraction, title generation, content moderation, and other catalog enrichment tasks._

**Trendyol-Vision-Master vs Flash:** Master is the larger expert VLM for critical or harder cases (especially category detection) and retains stronger general capabilities. For day-to-day, high-traffic catalog workloads on a single GPU, use [Trendyol/Trendyol-Vision-Flash](https://huggingface.co/Trendyol/Trendyol-Vision-Flash).

## Model Details

- **Architecture**: Qwen3.5-27B vision-language model (VLM; 27B parameters)
- **Base model**: [Qwen/Qwen3.5-27B](https://huggingface.co/Qwen/Qwen3.5-27B)
- **Training**: Full SFT on Trendyol catalog multimodal data
- **Languages**: Turkish (primary), English (secondary)
- **Modalities**: Image + text (multi-image supported)
- **Thinking mode**: Disabled at inference (`enable_thinking: False`)

## Intended Use

- Automate catalog moderation and enrichment tasks on Trendyol product listings.
- Classify products into categories from image + title by selecting from the category options provided in the prompt.
- Detect whether two product listings represent the same SKU.
- Extract structured attributes, brands, titles, and captions from product images.
- Moderate unsafe or policy-violating product content.
- Support research and evaluation use cases.

**Not intended for** general open-domain chat, medical/legal advice, surveillance, or any use described under Ethical Considerations. For low-latency single-GPU catalog enrichment without Master-tier category detection, use [Trendyol/Trendyol-Vision-Flash](https://huggingface.co/Trendyol/Trendyol-Vision-Flash).

## Supported Use Cases

Specialized for **e-commerce catalog operations** (moderation and enrichment). The tasks below are the primary, production-validated workloads; related catalog workflows can be prompted similarly, with best results on these patterns.

- **Category detection** — Choose the best matching category from the candidate list supplied in the prompt (image + title); the model returns the selected option index, not a free-form category name.
- **Product similarity** — Decide whether two listings (images and titles) refer to the same product.
- **Brand detection** — Infer the brand from product images with optional category context.
- **Attribute extraction** — Extract structured product attributes from image and title/description.
- **Title generation** — Produce a clean catalog title from the product image and a reference title.
- **Product caption** — Generate a grounded product description from the image and optional metadata.
- **Content safety classification** — Classify product image and title for catalog content safety (`0` = Forbidden, `1` = Fantasy, `2` = Safe). **Fantasy** means content that may be published but should be treated as adult (+18).
- **Per-pack quantity extraction** — Extract pack quantity and unit from image and long title.

See **Task Prompts** below for per-task prompt templates. All served requests use the same vLLM OpenAI chat payload shape; only the question text and image URLs change.

---

## Serving (vLLM)

Trendyol-Vision-Master can be served with vLLM as an OpenAI-compatible API. Multi-GPU serving is recommended for the 27B model (e.g. `--tensor-parallel-size 2`).

Validated with:

```bash
pip install "vllm==0.19.1"
```

```bash
vllm serve Trendyol/Trendyol-Vision-Master \
  --trust-remote-code \
  --max-model-len 16384 \
  --structured-outputs-config.backend xgrammar \
  --served-model-name Trendyol-Vision-Master \
  --interleave-mm-strings \
  --tensor-parallel-size 2
```

Every use case uses the same chat-completions payload template. Only the `messages` content (question text and `image_url` blocks) changes. Always disable thinking mode.

When calling the vLLM OpenAI API, use `image_url` content blocks:

```bash
curl http://localhost:8000/v1/chat/completions \
  -H "Content-Type: application/json" \
  -H "Authorization: Bearer EMPTY" \
  -d '{
    "model": "Trendyol-Vision-Master",
    "temperature": 0.0,
    "max_tokens": 16,
    "chat_template_kwargs": {"enable_thinking": false},
    "messages": [
      {
        "role": "user",
        "content": [
          {"type": "image_url", "image_url": {"url": "https://cdn.dsmcdn.com/ty103/product/media/images/20210412/17/79531657/163017575/1/1_org_thumb.jpg"}},
          {"type": "text", "text": "Ürün başlığı: 2'\''li Saten Robe Gecelik Sabahlık Şeftali Kısa Çeyiz Takımı 1013\n\nGörseli ve başlığı inceleyerek doğru kategori numarasını seç.\n0. Diğer\n1. Külot\n2. İç Çamaşırı Takımı\n3. Boxer\n4. Fantezi Külot\n5. Fantezi Sabahlık\n6. Fantezi Aksesuar\n7. Fantezi Sütyen\n8. Fantezi Jartiyer\n9. Fantezi Boxer\n10. Fantezi Kombinezon\n11. Fantezi Gecelik\n12. Spor Sütyeni\n13. Elbise\n14. Fantezi Slip\n15. Tunik\n16. Fantezi Atlet\n17. Büyük Beden Elbise\n18. Büyük Beden Fantezi Gecelik\n19. String\n20. Fantezi İç Çamaşır Takımı\n21. Gecelik\n22. Fantezi Ürünleri\n23. Alt - Üst Takım\n24. Fantezi Kostüm\n25. Pijama Takımı\n26. Büyük Beden Pijama Takımı\n27. Fantezi String\n28. Fantezi Babydoll\n29. Pijama Üstü\n30. Fantezi Çorap\n31. Sütyen\n32. Sabahlık\n33. Tesettür Elbise\n34. Büyük Beden Abiye Elbise\n35. Pijama Altı\n36. Bikini Üstü\n\nCevap olarak yalnızca numarayı yaz."}
        ]
      }
    ]
  }'
```

---

## Task Prompts

Pass any prompt below through the same vLLM payload template shown above. Replace `{placeholders}` with your data. Use `{"type": "image_url", "image_url": {"url": "..."}}` for each image (public HTTP(S) URL or data URL).

### 1. Category Detection

**Images:** 1 product image  
**Output:** category option index only (e.g. `21`)

```
Ürün başlığı: {title}

Görseli ve başlığı inceleyerek doğru kategori numarasını seç.
0. {option_0}
1. {option_1}
...
N. {option_n}

Cevap olarak yalnızca numarayı yaz.
```

### 2. Product Similarity

**Images:** before listing images, then after listing images  
**Output:** `1` (same SKU) or `0` (different)

```
E-ticaret kataloğundaki ürün benzerliği konusunda uzmansınız.
İki ürünün önceki ve sonraki başlık/görsellerini karşılaştırarak aynı ürün olup olmadığını belirle.
Ambalaj, arka plan veya model farkları tek başına fark sayılmaz; renk, boyut, miktar veya varyant farkları fark sayılır.
Sadece "1" veya "0" döndür.

Önceki Ürün Başlığı:
{before_title}
Önceki Ürün Resimleri:
[image_url...]
Sonraki Ürün Başlığı:
{after_title}
Sonraki Ürün Resimleri:
[image_url...]
```

Example payload fragment:

```python
content = [
    {"type": "text", "text": "E-ticaret kataloğundaki ürün benzerliği konusunda uzmansınız.\nİki ürünün önceki ve sonraki başlık/görsellerini karşılaştırarak aynı ürün olup olmadığını belirle.\nAmbalaj, arka plan veya model farkları tek başına fark sayılmaz; renk, boyut, miktar veya varyant farkları fark sayılır.\nSadece \"1\" veya \"0\" döndür.\n\nÖnceki Ürün Başlığı:\nBambu Kapaklı Vakumlu Borosilikat Yağdanlık | 3 Parça - 2lt\nÖnceki Ürün Resimleri:"},
    {"type": "image_url", "image_url": {"url": "https://cdn.dsmcdn.com/ty1757/prod/QC_ENRICHMENT/20250918/02/8683e8bc-61b1-3589-be35-3627320b56e9/1_org_thumb.jpg"}},
    {"type": "text", "text": "Sonraki Ürün Başlığı:\nBambu Kapaklı Vakumlu Borosilikat Yağdanlık Seti\nSonraki Ürün Resimleri:"},
    {"type": "image_url", "image_url": {"url": "https://cdn.dsmcdn.com/ty1804/prod/QC_ENRICHMENT/20251224/22/1e392dbd-51ab-3513-833a-87817b6b61b4/1_org_thumb.jpg"}},
    {"type": "text", "text": "Çıktı sadece \"1\" veya \"0\" olmalı. Açıklamayı pas geçin."},
]
# Expected output: "0"
```

### 3. Brand Detection

**Images:** product photos (1–16)  
**Output:** brand name or `Unknown`

```
Görsellerden, {category} kategorisinde yer alan ürünün markasını çıkar.
Sadece verilen görsellerde doğrulanabilen bilgilere dayan.
Emin olmadığında "Unknown" şeklinde cevap ver.
Sadece marka adını döndür.
```

Example:

```python
content = [
    {"type": "image_url", "image_url": {"url": "https://cdn.dsmcdn.com/mnresize/620/920/ty1571/prod/QC/20240924/23/a77c7933-c626-3753-9cad-fbb7388bff45/1_org_zoom.jpg"}},
    {"type": "image_url", "image_url": {"url": "https://cdn.dsmcdn.com/mnresize/620/920/ty1569/prod/QC/20240924/23/356395c3-c71c-37c5-88c4-5ef97899b3d8/1_org_zoom.jpg"}},
    {"type": "image_url", "image_url": {"url": "https://cdn.dsmcdn.com/mnresize/620/920/ty1571/prod/QC/20240924/23/112305b3-ba5d-3a28-8fd3-b1857749afca/1_org_zoom.jpg"}},
    {"type": "text", "text": "Görsellerden, Eldiven kategorisinde yer alan ürünün markasını çıkar.\nSadece verilen görsellerde doğrulanabilen bilgilere dayan.\nEmin olmadığında \"Unknown\" şeklinde cevap ver.\nSadece marka adını döndür."},
]
# Expected output: "Adidas"
```

### 4. Attribute Extraction

**Images:** 1 product image  
**Output:** JSON object

```
Bu görseldeki, başlığı '{title}' ve açıklaması '{description}' olan ürünün {attribute_list} bilgilerini json formatında çıkarır mısın?
```

Example:

```python
content = [
    {"type": "image_url", "image_url": {"url": "https://cdn.dsmcdn.com/ty1325/product/media/images/prod/QC/20240522/17/f63a7316-5998-3c7d-b45f-93ff8e9c34a2/1_org_zoom.jpg"}},
    {"type": "text", "text": "Bu görseldeki, başlığı 'Helen Bar Sandalyesi-mavi-9519q0119' ve açıklaması 'Maksimum kargolanma süresi: Sipariş tarihinden sonraki 7. gün Ortalama montaj süresi: 1 dakika Genişlik: 62cm Derinlik: 52cm Yükseklik: 96cm Oturak Yüksekliği: 44cm Ürün Ağırlığı: 10kg Kullanılan malzeme: Metal Kullanılan sünger: Yüksek yoğunluklu gri sünger Kullanılan kol: Metal Kollu Kullanılan ayak: Metal Ayaklı + Kromajlı Garanti süresi - Menşei: 24 Ay - Yerli Sipariş bazlı üretim-tedarik yapıldığından sipariş iptali yapılamamaktadır Anlaşmalı olunan ambar ve kargolarla bina kapısında teslimat yapılmaktadır' olan ürünün Garanti Süresi, Materyal, Model, Sandalye Kumaşı, Sandalye Sayısı, Tema / Stil bilgilerini json formatında çıkarır mısın?"},
]
# Expected output:
# {
#     "Garanti Süresi": "2 Yıl",
#     "Materyal": "Metal",
#     "Model": "Bar Sandalyesi",
#     "Sandalye Kumaşı": "çıkarılamadı",
#     "Sandalye Sayısı": "1",
#     "Tema / Stil": "Modern"
# }
```

### 5. Title Generation

**Images:** 1 product image  
**Output:** plain-text title

```
Ürün fotoğrafı ile '{reference_title}' bilgisini karşılaştırıp, Trendyol katalog moderasyon kurallarına göre yanıltıcı ifadelerden kaçınarak net bir başlık üret; çıktıyı düz metin ver.
```

Example:

```python
content = [
    {"type": "image_url", "image_url": {"url": "https://cdn.dsmcdn.com/ty1684/prod/QC/20250528/12/51036511-2e99-3a97-a594-e783d8c7bb8e/1_org_zoom.jpg"}},
    {"type": "text", "text": "Ürün fotoğrafı ile '%100 Polarize UV400 Korumalı Füme Renk Orijinal Oval Erkek Güneş Gözlüğü DKG6273C3' bilgisini karşılaştırıp, Trendyol katalog moderasyon kurallarına göre yanıltıcı ifadelerden kaçınarak net bir başlık üret; çıktıyı düz metin ver."},
]
# Expected output: Polarize UV 400 korumalı füme lensli siyah dikdörtgen güneş gözlüğü
```

### 6. Product Caption

**Images:** 1 product image  
**Output:** English product description

```
Without speculating about details you cannot see, describe this product based on the image and the information provided.
Product title: {title}
Brand: {brand}
First decide which object is the product, review OCR for brand/model/title clues, then analyze colors, shape, material, pattern, and other grounded details.
```

Example:

```python
content = [
    {"type": "image_url", "image_url": {"url": "https://cdn.dsmcdn.com/ty1742/prod/QC_PREP/20250831/14/e1be2979-cfe6-3a32-b7d8-6e5152722b20/1_org_zoom.jpg"}},
    {"type": "text", "text": "Without speculating about details you cannot see, describe this product based on the image and the information provided.\nProduct title: çok renkli çiçek desenli anne elbisesi\nBrand: mihvera\nFirst decide which object is the product, review OCR for brand/model/title clues, then analyze colors, shape, material, pattern, and other grounded details."},
]
# Expected output: viscose floral print maxi dress, gathered round neckline, long sleeves, relaxed loose fit, floor length hem, multicolor design on dark base
```

### 7. Content Safety Classification

**Images:** 1 product image  
**Labels:** `0` (Forbidden) = not publishable, `1` (Fantasy) = publishable but adult (+18) content, `2` (Safe) = publishable without restriction.

```
Ürün Başlığı: {title}

Bu ürün görselini ve başlığını inceleyerek moderasyon sınıflandırması yap. Sonucu 0 (Forbidden), 1 (Fantasy) veya 2 (Safe) olarak ver.
```

Example:

```python
content = [
    {"type": "image_url", "image_url": {"url": "https://cdn.dsmcdn.com/ty1624/prod/QC/20250117/10/46b6e784-4674-3ae6-a4c0-55a19b83353a/1_org_zoom.jpg"}},
    {"type": "text", "text": "Ürün Başlığı: 8. SINIF LGS - MEB BÖYLE SORAR SARMAL BRANŞ DENEME SETİ İKİNCİ DOZ (2025-LGS)\n\nBu ürün görselini ve başlığını inceleyerek moderasyon sınıflandırması yap. Sonucu 0 (Forbidden), 1 (Fantasy) veya 2 (Safe) olarak ver."},
]
# Expected output: "2"
```

### 8. Per-Pack Quantity Extraction

**Images:** 1 product image  
**Output:** `{"amount": <number|null>, "unit": "KG"|"L"|"PIECE"|null}`

```
Extract total product quantity from the title and image. Title is the primary source; use the image only when the title is missing or ambiguous.
Return only JSON: {"amount": <number|null>, "unit": "KG"|"L"|"PIECE"|null}
Convert g→KG, ml/cc→L, counts→PIECE. Ignore model numbers, storage, wattage, and dimensions. If unclear, return {"amount": null, "unit": null}.

Product Title: {title}
```

Example:

```python
content = [
    {"type": "text", "text": "Extract total product quantity from the title and image. Title is the primary source; use the image only when the title is missing or ambiguous.\nReturn only JSON: {\"amount\": <number|null>, \"unit\": \"KG\"|\"L\"|\"PIECE\"|null}\nConvert g→KG, ml/cc→L, counts→PIECE. Ignore model numbers, storage, wattage, and dimensions. If unclear, return {\"amount\": null, \"unit\": null}.\n\nProduct Title: Seda 6'lı Çay Bardağı"},
    {"type": "image_url", "image_url": {"url": "https://cdn.dsmcdn.com/ty1791/prod/QC_ENRICHMENT/20251121/11/be96b90c-9373-33be-bdc2-22c0e794c0f6/1_org_zoom.jpg"}},
]
# Expected output: {"amount": 6, "unit": "PIECE"}
```

---

## Limitations

- **Domain specificity**: Optimized for Trendyol e-commerce product images and Turkish catalog text; may not generalize to other domains.
- **Category scope**: Category options are provided in the prompt; out-of-scope products should be mapped to an option such as `"Diğer"` in that list.
- **GPU requirements**: 27B model requires multi-GPU serving or high-VRAM hardware.
- **Language bias**: Turkish prompts and outputs are primary; English performance varies by task.
- **Not a general assistant**: Fine-tuned for structured catalog tasks, not open-ended conversation.

## Ethical Considerations

- Designed for catalog quality and moderation workflows.
- Ensure compliance with data protection regulations when processing product and user-generated content.
- Monitor for biased decisions across product categories and brands.
- Not intended for surveillance, misinformation, or harmful content generation.
- Safety and moderation labels reflect training for catalog workflows and may not match every jurisdiction or platform policy; human review is recommended for high-impact decisions.

## Citation

```bibtex
@misc{trendyol-vision-master,
  title={Trendyol-Vision-Master: Catalog Quality Vision-Language Model (VLM)},
  author={Trendyol Data Science Team},
  year={2026},
  howpublished={\url{https://huggingface.co/Trendyol/Trendyol-Vision-Master}}
}
```

```bibtex
@misc{qwen3.5,
  title={{Qwen3.5}: Towards Native Multimodal Agents},
  author={{Qwen Team}},
  month={February},
  year={2026},
  url={https://huggingface.co/Qwen/Qwen3.5-27B}
}
```

## Model Card Authors

- Trendyol Data Science Team

## License

This model is licensed under the [Creative Commons Attribution 4.0 International License (CC BY 4.0)](https://creativecommons.org/licenses/by/4.0/).

You are free to share and adapt the model for any purpose, even commercially, as long as you give appropriate credit and indicate if changes were made.

This release is a fine-tune of [Qwen/Qwen3.5-27B](https://huggingface.co/Qwen/Qwen3.5-27B), which is licensed under [Apache License 2.0](https://www.apache.org/licenses/LICENSE-2.0). Redistribution of this derivative continues to satisfy Apache-2.0 notice and attribution requirements for the base model; retain the Apache-2.0 license text and any NOTICE attributions shipped with Qwen3.5-27B when redistributing.

For the full CC BY 4.0 license text, see: https://creativecommons.org/licenses/by/4.0/legalcode

---

_Released by the Trendyol Data Science Team._