ozertuu/turkish-food-contrastive-triplets

🥘 Turkish & Multi-Lingual Food Contrastive Triplets Dataset

This dataset contains 30,001 culinary triplets (Anchor, Positive, Hard Negative) designed for training contrastive dense retrieval models, embedding projection heads, and fine-grained food similarity metrics.

Used for training the ozertuu/bge-m3-food-projection-head.

---

📊 Dataset Structure

Each record consists of:

• anchor (string): The base query or canonical food name (e.g., "Kuru Fasulye").
• positive (string): A gastronomically aligned variation, preparation, or title match (e.g., "Soslu etli kuru fasulye yemeği").
• negative (string): A hard negative containing lexical overlap but representing an unrequested ingredient, raw item, or conflicting dish (e.g., "Kuru fasulye konservesi").
• type (string): The triplet generation contrast strategy (cooking_method_contrast, category_alignment, brand_demotion, raw_ingredient_demotion).

Sample Data

{
  "anchor": "Fried Eggs",
  "positive": "Yeşil Biber Kızartması",
  "negative": "Beyaz peynirli menemen",
  "type": "cooking_method_contrast"
}

---

🚀 How to Load

from datasets import load_dataset

dataset = load_dataset("ozertuu/turkish-food-contrastive-triplets")
print(dataset["train"][0])