metadata

title: Recipe-GPT
emoji: 🍳
colorFrom: yellow
colorTo: red
sdk: gradio
sdk_version: 4.27.0
app_file: app.py
pinned: false
license: mit
tags:
  - gpt2
  - gradio
  - transformers
  - recipe-generation
  - huggingface
short_description: 'Turn ingredients into step-by-step cooking directions. '

Recipe-GPT

This project fine-tunes the pretrained openai-community/gpt2 model using Hugging Face’s transformers and datasets libraries on a large corpus of human-written recipes.

The goal is to teach GPT-2 how to generate coherent cooking instructions from a list of ingredients.

Key Features:

Base Model: GPT-2 (117M) from Hugging Face Model Hub
Frameworks:
- transformers for model loading, training, and generation
- datasets for loading and managing large recipe data
- Trainer API for end-to-end fine-tuning
Custom Tokens:
- Added <start> and <end> tokens to mark recipe boundaries
- Tokenizer and model embeddings resized accordingly
Data Format:
- Recipes formatted as plain text blocks with titles, ingredients, and step-by-step directions
Training Strategy:
- Causal language modeling (not masked)
- Evaluated on validation set each epoch
- Supports CPU, CUDA, and Apple MPS backends

This setup allows the model to learn full-text generation patterns and structure, making it effective at translating structured ingredient lists into complete, human-readable cooking instructions.

Dataset

Source: tengomucho/all-recipes-split

2.1M+ recipes with:

title
ingredients
directions

How to Use

Install dependencies

pip install -r requirements.txt

Download dataset and tokenize

python scripts/dataset.py

Train the model

python scripts/train.py

Run Inference

python scripts/inference.py

Enter comma-separated ingredients when prompted, for example,

Enter ingredients (comma-separated): chicken, rice, garlic