C-three-AN-Nourich/Recipe-1m-processed

Recipe1M Processed Dataset Collection

Abstract

This repository provides two complementary processed datasets derived from the Recipe1M dataset:

1. Instruction-Ingredient Alignment Dataset (processed_instructions.json): Fine-grained instruction-ingredient alignment at the fragment level with multi-granularity ingredient representations
2. Multimodal Recipe Dataset (train_data.json, val_data.json, test_data.json): Recipe data with ingredient-image mappings and recipe images for multimodal learning

Together, these datasets enable comprehensive recipe understanding tasks including ingredient recognition, instruction generation, multimodal learning, and cooking process modeling.

---

---

Dataset Overview

Dataset	Files	Key Features	Primary Use Cases
Instruction-Ingredient Alignment	processed_instructions.json	Fragment-level alignment, multi-granularity ingredients	Instruction parsing, ingredient extraction, process understanding
Multimodal Recipe Data	train_data.json, val_data.json, test_data.json	Image mappings, standardized ingredients, train/val/test splits	Multimodal learning, image-based retrieval, model training

---

Dataset 1: Instruction-Ingredient Alignment

Overview

Fine-grained instruction-ingredient alignment where each cooking instruction is fragmented at ingredient boundaries and aligned with corresponding ingredient information at three levels of granularity.

File Structure

processed_instructions.json

• Format: JSON dictionary with recipe IDs as keys
• Recipe ID Format: 10-character hexadecimal (e.g., e40cfb8510)

Data Schema

{
  "id": "recipe_id",
  "instructions": [
    {
      "instruction": "cooking step text fragment",
      "raw_ing": "original ingredient with quantity",
      "final_ing": "standardized ingredient name",
      "no_with_ing": "ingredient without modifiers"
    }
  ]
}

Field Descriptions

Top-Level Fields

Field	Type	Description	Example
id	String	Unique recipe identifier	"e40cfb8510"
instructions	List	Cooking steps with ingredient alignment	See below

Instruction Object Fields

Field	Type	Description	Example
instruction	String	Instruction text fragment	"stir together flour,"
raw_ing	String/null	Original ingredient with quantities	"2 c. flour"
final_ing	String/null	Standardized ingredient name	"flour"
no_with_ing	String/null	Ingredient without modifiers	"flour"

Note: null values indicate instruction steps without specific ingredient references (e.g., cooking techniques, equipment preparation).

Example Data

{
  "id": "e40cfb8510",
  "instructions": [
    {
      "instruction": "stir together flour,",
      "raw_ing": "2 c. flour",
      "final_ing": "flour",
      "no_with_ing": "flour"
    },
    {
      "instruction": " soda ",
      "raw_ing": "1 teaspoon baking soda",
      "final_ing": "baking soda",
      "no_with_ing": "baking soda"
    },
    {
      "instruction": "and salt.",
      "raw_ing": "1/2 teaspoon salt",
      "final_ing": "salt",
      "no_with_ing": "salt"
    },
    {
      "instruction": "stir into a very stiff batter.",
      "raw_ing": null,
      "final_ing": null,
      "no_with_ing": null
    },
    {
      "instruction": "bake 1 hour at 325 degrees.",
      "raw_ing": null,
      "final_ing": null,
      "no_with_ing": null
    }
  ]
}

Key Features

• Instruction Fragmentation: Instructions split at ingredient boundaries while maintaining grammatical structure
• Multi-Granularity: Three levels of ingredient representation (raw, final, no modifiers)
• Complete Process: Includes all steps, both ingredient-specific and technique-only
• Ingredient Repetition: Tracks multiple uses of the same ingredient throughout the recipe

---

Dataset 2: Multimodal Recipe Data

Overview

Structured recipe data with ingredient-image mappings, standardized ingredient names, and associated recipe images. Split into training, validation, and test sets for machine learning workflows.

File Structure

File	Purpose	Recipe IDs Example
train_data.json	Training set	df203c7b00, da3722f92a, ...
val_data.json	Validation set	da3722f92a, ...
test_data.json	Test set	da36dcb1f9, ...

Data Schema

{
  "recipe_id": {
    "ing_img_ids": [image_id or null, ...],
    "ing_text_18k": ["standardized_name" or null, ...],
    "ing_texts": ["original ingredient text" or null, ...],
    "instructions": ["step 1", "step 2", ...],
    "recipe_img_ids": ["image1.jpg", "image2.jpg", ...]
  }
}

Field Descriptions

Field	Type	Description	Example
ing_img_ids	List[int/null]	Image IDs from ingredient database	[17263, 8861, None, ...]
ing_text_18k	List[str/null]	Standardized names (18K vocabulary)	['butter', 'olive oil', None, ...]
ing_texts	List[str/null]	Original text with quantities	['1 teaspoon butter', '2 teaspoons olive oil', ...]
instructions	List[str]	Cooking steps in order	['melt butter', 'add garlic', ...]
recipe_img_ids	List[str]	Recipe image filenames	['50cbeb2173.jpg', '95f9aa7309.jpg', ...]

Example Data

{
  "da3722f92a": {
    "ing_img_ids": [17263, 8861, 11498, 3496, None, 578],
    "ing_text_18k": ["butter", "olive oil", "garlic cloves", "shrimp", None, "lemon"],
    "ing_texts": [
      "1 teaspoon butter",
      "2 teaspoons olive oil",
      "3 garlic cloves, minced or pressed",
      "1 lb california shrimp or 1 lb spiny lobster",
      None,
      "1 lemon, juice of"
    ],
    "instructions": [
      "melt butter and oil together in saute pan.",
      "add garlic, saute for one minute, and add shrimp.",
      "saute for one minute, add wine, lemon juice, salt, and pepper.",
      "saute quickly while sauce reduces and shrimp turns pink.",
      "do not overcook.",
      "sprinkle with parsley before serving."
    ],
    "recipe_img_ids": [
      "50cbeb2173.jpg",
      "95f9aa7309.jpg",
      "b81c7c2b13.jpg",
      "977929a1ff.jpg"
    ]
  }
}

Key Features

• Multimodal Alignment: Links text ingredients to visual representations
• Standardization: 18K ingredient vocabulary for consistency
• Image References: Both ingredient images and recipe images
• Train/Val/Test Split: Ready for supervised learning workflows
• Null Handling: Graceful handling of missing image mappings

---

---

Related Resources

These processed datasets build upon:

• Recipe1M Dataset: Original source of recipe data and images
• Ingredient Image Database: Referenced by ing_img_ids
• Recipe Image Database: Referenced by recipe_img_ids
• 18K Ingredient Vocabulary: Used for standardization

---

License

Please refer to the original Recipe1M dataset license and terms of use. These processed versions maintain the same licensing requirements as the source dataset.

---

Citation

If you use these datasets, please cite the original Recipe1M work:

@inproceedings{salvador2017learning,
  title={Learning Cross-Modal Embeddings for Cooking Recipes and Food Images},
  author={Salvador, Amaia and Hynes, Nicholas and Aytar, Yusuf and Marin, Javier and 
          Ofli, Ferda and Weber, Ingmar and Torralba, Antonio},
  booktitle={Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition},
  year={2017}
}

---

Acknowledgments

These datasets are derived from the Recipe1M dataset. We acknowledge the original authors for their valuable contribution to the research community. The processing includes instruction-ingredient alignment and multimodal mappings to enable new research directions.

---

Appendix: Quick Reference

Dataset 1: Instruction-Ingredient Alignment

• Files: processed_instructions.json
• Key Features: Fragment-level alignment, 3 granularity levels
• Best For: Instruction parsing, ingredient extraction

Dataset 2: Multimodal Recipe Data

• Files: train_data.json, val_data.json, test_data.json
• Key Features: Image mappings, train/val/test split
• Best For: Multimodal learning, ingredient recognition

Common Recipe ID Format

• Format: 10-character hexadecimal
• Example: da3722f92a, e40cfb8510
• Usage: Links recipes across both datasets