GLM-5.3 is live. Run Z.ai's latest model on Friendli Model APIs. Try it today ➜
Turhan123
Available on FriendliAI
Run this model inference on single tenant GPU with unmatched speed and reliability at scale.
Run this model inference with full control and performance in your environment.
Model Details
Model Provider
Turhan123
Model Tree
Qwen/Qwen2-VL-2B-Instruct
Input Modalities
Output Modalities
Supported Functionality
This study was conducted to address the challenge of global VLM models failing to recognize local cultural elements, specifically Turkish cuisine.
General-purpose VLMs often misclassify specific cultural dishes (e.g., confusing Mantı with Ravioli or Pasta). This model was developed to identify Turkish meals and answer nutrition questions about them.
It utilizes QLoRA (LoRA adapters trained on a 4-bit quantized base model) for efficient fine-tuning, preserving the capabilities of the base Qwen2-VL model while injecting domain-specific knowledge about Turkish cuisine.
The model has been trained to recognize and provide nutritional information for the following Turkish dishes:
Adana Kebap, Patlıcan Kebabı, İskender, Taş Kebabı, Et Döner, Hünkar Beğendi, Mantı, Hamsi Tava
Beyaz Lahana Sarması, Biber Dolması, Midye Dolma, Mumbar Dolması, Yaprak Sarma
Gözleme, Su Böreği, Lahmacun, Kıymalı Pide, Simit
Çiğ Köfte, İçli Köfte, Çoban Salatası, Kısır, Midye Tava, Mücver, Kabak Mücver, Tantuni, Sucuklu Yumurta, Cacık, Menemen
Mercimek Çorbası, Yayla Çorbası, Domates Çorbası, Şehriye Çorbası, Tarhana Çorbası
Kuru Fasulye, Taze Fasulye
Mercimek Köftesi
Baklava, Sütlaç, Kazandibi, Tulumba Tatlısı, Lokma, Lokum, Kemal Paşa Tatlısı, Kalburabastı
Türk Kahvesi, Sahlep
Note: The model may attempt to recognize other Turkish foods, but it was only trained on the 47 dishes listed above, and answers for other dishes are unreliable.
The model is intended for:
You can use this model with transformers and peft.
bash
pip install -U transformers peft accelerate pillowpython
import torchfrom PIL import Imagefrom transformers import Qwen2VLForConditionalGeneration, Qwen2VLProcessorfrom peft import PeftModel # 1. Load Base Model and Processorbase_model_id = "Qwen/Qwen2-VL-2B-Instruct"adapter_id = "Turhan123/turkish-cuisine-vlm" # Load Base Model (fp16 works on T4 and newer GPUs)model = Qwen2VLForConditionalGeneration.from_pretrained( base_model_id, torch_dtype=torch.float16, device_map="auto",)processor = Qwen2VLProcessor.from_pretrained(base_model_id) # 2. Load the Fine-Tuned Adaptermodel = PeftModel.from_pretrained(model, adapter_id) # System prompt used during trainingSYSTEM_PROMPT = "Sen Türk mutfağı konusunda uzman bir diyetisyen ve VLM asistanısın. Yemekleri tanı ve besin değerlerini söyle." # 3. Define Helper Functiondef ask_dietitian(image_path, question="Bu yemek nedir ve besin değerleri nasıldır?"): image = Image.open(image_path) conversation = [ {"role": "system", "content": [{"type": "text", "text": SYSTEM_PROMPT}]}, { "role": "user", "content": [ {"type": "image", "image": image}, {"type": "text", "text": question}, ], }, ] text_prompt = processor.apply_chat_template(conversation, add_generation_prompt=True) inputs = processor( text=[text_prompt], images=[image], padding=True, return_tensors="pt" ).to(model.device) output_ids = model.generate(**inputs, max_new_tokens=150, do_sample=False) generated_ids = [ output_ids[len(input_ids):] for input_ids, output_ids in zip(inputs.input_ids, output_ids) ] return processor.batch_decode(generated_ids, skip_special_tokens=True)[0] # Example Usage# response = ask_dietitian("kebap.jpg")# print(response)The model was trained on a Custom Turkish Food Dataset curated specifically for this project.
Example (baklava):
Q: Kilo aldırır mı? A: Evet, şerbet (şeker) ve yağ içeriği çok yüksek olduğu için porsiyon kontrolü yapılmazsa hızla kilo aldırır.
transformers, peft, bitsandbytes and trl (SFTTrainer).Optimization Hyperparameters:
q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_projNo quantitative benchmark was run. The model was checked informally with a "blind test" on random images collected from the internet (not seen during training).
Observations:
I would like to express my special thanks to the creators of the following resources, which made this project possible:
Special Thanks for Tutorial & Inspiration:
Dataset Sources:
Technical References:
pip install -U transformers peft accelerate pillowimport torchfrom PIL import Imagefrom transformers import Qwen2VLForConditionalGeneration, Qwen2VLProcessorfrom peft import PeftModel # 1. Load Base Model and Processorbase_model_id = "Qwen/Qwen2-VL-2B-Instruct"adapter_id = "Turhan123/turkish-cuisine-vlm" # Load Base Model (fp16 works on T4 and newer GPUs)model = Qwen2VLForConditionalGeneration.from_pretrained( base_model_id, torch_dtype=torch.float16, device_map="auto",)processor = Qwen2VLProcessor.from_pretrained(base_model_id) # 2. Load the Fine-Tuned Adaptermodel = PeftModel.from_pretrained(model, adapter_id) # System prompt used during trainingSYSTEM_PROMPT = "Sen Türk mutfağı konusunda uzman bir diyetisyen ve VLM asistanısın. Yemekleri tanı ve besin değerlerini söyle." # 3. Define Helper Functiondef ask_dietitian(image_path, question="Bu yemek nedir ve besin değerleri nasıldır?"): image = Image.open(image_path) conversation = [ {"role": "system", "content": [{"type": "text", "text": SYSTEM_PROMPT}]}, { "role": "user", "content": [ {"type": "image", "image": image}, {"type": "text", "text": question}, ], }, ] text_prompt = processor.apply_chat_template(conversation, add_generation_prompt=True) inputs = processor( text=[text_prompt], images=[image], padding=True, return_tensors="pt" ).to(model.device) output_ids = model.generate(**inputs, max_new_tokens=150, do_sample=False) generated_ids = [ output_ids[len(input_ids):] for input_ids, output_ids in zip(inputs.input_ids, output_ids) ] return processor.batch_decode(generated_ids, skip_special_tokens=True)[0] # Example Usage# response = ask_dietitian("kebap.jpg")# print(response)