This is basically a fork from https://github.com/Glorf/recipenlg. Please refer to the original repo for more details and to use as reference.
The objective of this project is to generate high-quality, context-aware cooking recipes by exploring and using the latest advancements in natural language processing. The project is targeted towards users who are interested in cooking and want help in generating recipes based on a list of ingredients as an input.
- reorganize the code to have a better structure, it sucks
- add the sequence length of the tokenized dataset in the h5 file name.
- finetune llama3.1 8b using some kind of lora.
- implement some RAG in there
- remove temporary evaluation things and move to config_evaluation.yaml
- add the possibility to evaluate a model with a specific checkpoint
- make this for run_evaluation.py and its pipeline
You need Miniconda or Virtualenv to run this project. We recommend using Miniconda. We are using Python 3.12.
conda create -n recipe_generation python=3.12
First install PyTorch 2.4 for your device: https://pytorch.org/get-started/locally/. We recommend using the version of cuda your device supports.
Afterwards, you can install the rest of the librairies.
For locked versions of the libraries, you can use the following command:
conda install -c conda-forge --yes --file requirements.txt
To try newer versions of the libraries, you can use the following command:
conda install -c conda-forge --yes --file requirements-dev.txt
We also need to download the full dataset that you will find here: Full dataset. The code will preprocess this data during the finetuning phase. We do not use Git LFS yet.
The default config /config/config_finetuning.yaml will first create the
necessary datasets (requires 10GB of space, 45 minutes on an i7 7700k or 15
minutes on a ryzen 9 7950X3D) for the GPT2 (or specified) model. Afterwards,
this will start the fine-tuning of
the specified model in the config (gpt2, opt, llama).
python main_finetuning.py
The default config /config/config_generation.yaml will create recipes from the
GPT2 (or specified) model. You can set the prompt and the number of times you
want the generation to repeat for a specific prompt. This is also used to create
a fine-tuned dataset of recipes which is used in the evaluation pipeline. We
might add a link to a trained model to download to test the generation without
pretraining locally.
python run_generation.py
The default config /config/config_generation.yaml has a evaluate flag which
will evaluate the model on the test set. This will generate recipes and evaluate
them using various metrics. The results will be saved in the results folder.
If you use the RecipeNLG dataset and this code or the original code, use the following BibTeX entry since the work is closely related to the following paper:
@inproceedings{bien-etal-2020-recipenlg,
title = "{R}ecipe{NLG}: A Cooking Recipes Dataset for Semi-Structured Text Generation",
author = "Bie{\'n}, Micha{\l} and
Gilski, Micha{\l} and
Maciejewska, Martyna and
Taisner, Wojciech and
Wisniewski, Dawid and
Lawrynowicz, Agnieszka",
booktitle = "Proceedings of the 13th International Conference on Natural Language Generation",
month = dec,
year = "2020",
address = "Dublin, Ireland",
publisher = "Association for Computational Linguistics",
url = "https://www.aclweb.org/anthology/2020.inlg-1.4",
pages = "22--28",
}