docs: fill in dataset card documentation (description, biases, limitations, curation rationale)

README.md

@@ -14,7 +14,123 @@ configs:
   data_files:
   - split: train
     path: data/train-*
+language:
+- en
+license: other
+multilinguality:
+- monolingual
+size_categories:
+- 1M<n<10M
+pretty_name: All Recipes
+task_categories:
+- text2text-generation
+- text-generation
 ---
+
 # Dataset Card for "all-recipes"
 
-[More Information needed](https://github.com/huggingface/datasets/blob/main/CONTRIBUTING.md#how-to-contribute-to-the-dataset-cards)
+## Dataset Description
+
+- **Homepage:** [cookbooks.com](https://www.cookbooks.com)
+- **Repository:** [corbt/all-recipes](https://huggingface.co/datasets/corbt/all-recipes)
+- **Paper:** N/A
+- **Point of Contact:** [More Information Needed]
+
+### Dataset Summary
+
+`all-recipes` is a collection of 2,147,248 cooking recipes scraped from cookbooks.com and reformatted as input/output pairs for instruction fine-tuning of language models. Each example contains a recipe name and ingredient list as the `input` field, with the preparation steps as the expected output. The dataset is intended for fine-tuning generative models on recipe text generation tasks.
+
+### Languages
+
+English only. Measurements and terminology follow US customary conventions.
+
+## Dataset Structure
+
+### Data Instances
+
+Each example contains a single `input` field formatted as:
+
+```
+Recipe: <recipe title>
+Ingredients: <ingredient list>
+```
+
+The expected output (preparation steps) is not stored as a separate field; the dataset is structured for use with instruction-following fine-tuning frameworks where the model generates the directions from the input prompt.
+
+### Data Fields
+
+- `input` (`str`): A formatted string combining the recipe title and ingredient list, suitable for use as a prompt in instruction fine-tuning.
+
+### Data Splits
+
+The dataset contains a single `train` split with 2,147,248 examples.
+
+## Dataset Creation
+
+### Curation Rationale
+
+The dataset was assembled to provide a large-scale instruction fine-tuning corpus for recipe generation. cookbooks.com hosts a large public archive of user-submitted recipes, making it a convenient single-source scrape for culinary text at scale. The input/output format was chosen to match the prompt structure used in instruction-tuned language model training (e.g., Alpaca-style fine-tuning), where a model learns to produce directions given a recipe name and ingredient list.
+
+### Source Data
+
+All 2,147,248 recipes derive from a scrape of cookbooks.com, a US-based recipe hosting platform with user-contributed content. No additional sources were incorporated. No manual curation or editorial filtering was applied beyond deduplication.
+
+#### Initial Data Collection and Normalization
+
+Recipes were collected via automated scraping of cookbooks.com. The raw scraped data was reformatted into the `input` field structure. No normalization of ingredient names, measurement units, or recipe titles was performed.
+
+#### Who are the source language producers?
+
+Recipes were originally authored by cookbooks.com users, a predominantly US-based, English-speaking community.
+
+### Annotations
+
+The dataset contains no human annotations. The `input` field structure is a programmatic reformatting of the scraped recipe data.
+
+### Personal and Sensitive Information
+
+Recipe authors on cookbooks.com may have associated usernames or attribution text. This dataset does not retain author attribution fields; the `input` field contains only recipe title and ingredient text.
+
+## Considerations for Using the Data
+
+### Social Impact of Dataset
+
+Models fine-tuned on this dataset acquire culinary text generation capabilities that can be used in recipe generation applications, cooking assistants, and meal planning tools. The dataset's distributional characteristics (described below) will be reflected in model outputs unless researchers apply targeted fine-tuning or retrieval augmentation to address coverage gaps.
+
+### Discussion of Biases
+
+**Geographic and cultural skew.** The dataset derives entirely from cookbooks.com, a US-based platform with a user base concentrated in North America. Cuisines from Africa, South Asia, Southeast Asia, Latin America, the Caribbean, the Middle East, East Asia, and indigenous food traditions are substantially underrepresented relative to their global prevalence. Models trained on this dataset will generate recipes that skew toward mainstream North American and European cooking styles.
+
+**Measurement and unit conventions.** All ingredient quantities use US customary units (cups, tablespoons, teaspoons, ounces, pounds, Fahrenheit). Models fine-tuned on this corpus will default to these conventions and are unlikely to produce metric-unit output without additional prompting or fine-tuning.
+
+**Dietary category distribution.** The dataset contains no structured dietary labels (vegan, vegetarian, halal, kosher, allergen-free, gluten-free). cookbooks.com's user-submitted content skews toward conventional Western cooking, so animal-product recipes predominate in the corpus. Models trained on this dataset will reflect this distribution and may underperform on plant-based or restricted-diet recipe generation without targeted fine-tuning.
+
+**No quality filtering.** Recipes are user-submitted and have not been editorially reviewed. Recipe quality, clarity of instructions, and accuracy of ingredient quantities vary substantially across the corpus.
+
+### Other Known Limitations
+
+**English only.** The dataset is entirely in English and is not suitable for multilingual recipe generation.
+
+**No nutritional metadata.** The dataset does not include caloric content, macronutrient breakdown, allergen information, or serving size normalization.
+
+**No structured output field.** The preparation steps (directions) are embedded in the expected completion rather than stored as a structured field, which limits the dataset's utility for tasks requiring structured recipe parsing (e.g., ingredient extraction, step segmentation) without additional preprocessing.
+
+**License ambiguity.** The dataset derives from user-submitted content on cookbooks.com. The terms of use for the source platform should be reviewed before deploying models trained on this data in commercial applications.
+
+## Additional Information
+
+### Dataset Curators
+
+Dataset compiled by [corbt](https://huggingface.co/corbt).
+
+### Licensing Information
+
+The source content was scraped from cookbooks.com under that platform's terms of use. Researchers should review cookbooks.com's terms of service before using this dataset in commercial applications. No explicit open license has been assigned to this dataset.
+
+### Citation Information
+
+No paper is associated with this dataset. If you use it in published research, please cite the source platform and the dataset card:
+
+```
+corbt. (2023). all-recipes [Dataset]. Hugging Face. https://huggingface.co/datasets/corbt/all-recipes
+```