This program uses cooking datasets, such as those from Cookpad, to generate ingredients and instructions from a recipe title.
pip install -r requirements.txtPlease modify the installation of PyTorch according to your OS and computing environment. For more details, please refer to this site.
If you would like to use Flash Attention, please refer to this link and follow the instructions for installation.
For easy fine-tuning, you can use the Jupyter Notebook provided in the examples folder.
| Dataset Name | Language | Train Dataset Size | Validation Dataset Size | Test Dataset Size | All Dataset Size | URL | Seed |
|---|---|---|---|---|---|---|---|
| Cookpad dataset (Recipe data) | Japanese | https://www.nii.ac.jp/dsc/idr/cookpad/ | |||||
| data_recipes_instructor | English | https://huggingface.co/datasets/Erik/data_recipes_instructor | |||||
| llama2-TR-recipe | Turkish | https://huggingface.co/datasets/mertbozkurt/llama2-TR-recipe | |||||
| Recipes_Greek | Greek | https://huggingface.co/datasets/Depie/Recipes_Greek | |||||
| all-recipes-sm | English | https://huggingface.co/datasets/AWeirdDev/all-recipes-sm | |||||
| zh-tw-recipes-sm | Chinese | https://huggingface.co/datasets/AWeirdDev/zh-tw-recipes-sm | |||||
| all-recipes-xs | English | https://huggingface.co/datasets/AWeirdDev/all-recipes-xs | |||||
| aya-telugu-food-recipes | Telugu | https://huggingface.co/datasets/SuryaKrishna02/aya-telugu-food-recipes | |||||
| thai_food_v1.0 | Thai | https://huggingface.co/datasets/pythainlp/thai_food_v1.0 |
Please save the obtained Cookpad dataset in the data folder.
The CSV file of the Cookpad dataset contains the following information.
id,title,steps,ingredients
ad7d585b06850f8437ff5fb97d3c7a823ff21bb1,豚の角煮,鍋に、水とたっぷりのお酒、ねぎの使わない葉の部分、しょうがの皮、にんにくを入れて、２，３時間煮込みます。その間、あくや浮いてきた脂を丁寧に取りましょう。煮込んだお肉を、いったん水で洗いましょう。鍋に、豚肉をいれて、酒、砂糖、みりん、醤油、しょうが（薄切り）、にんにくで煮込みます。落とし蓋をして１時間。食べるちょっと前にねぎを入れて、味がついたらたべましょう。写真のは、ちんげん菜を入れてみました。,"しょうが(お好みで),ニンニク(お好みで),ねぎ(１本),豚肉(バラのブロック２パック),砂糖(小さじ１から２くらい),酒(たくさん（安い日本酒でいい）),醤油(適量（味見しながらね）),みりん(大さじ３くらい)"
4afce5687dc173ad4fef943b686582a1cd06e264,スペシャルピーマンの肉詰め,にんじんとれんこんをおろし金でおろします。挽肉と玉ねぎのみじん切りを加えよく塩コショウを加え、ピーマンに詰め、あとは焼くだけ。少し蒸らして火を通しできあがり。たれはおろしだれが一番！,"にんじん(2本),ピーマン(4つ),れんこん(小1),豚肉(挽肉 250g),おろしだれ(),コショウ(少々),塩(少々),たまねぎ(１つ)"
030833ed4e8dab3aa1e9d75edc1681efb368434f,簡単チーズリゾット,米は研がずにそのまま使っちゃえ。水も適当に。普通炊くよりも多めです。火加減は、はじめ強め、沸騰してきたらブイヨンをお米のかたさを見ながら…、水分が足りなくてリゾットができたら、取り皿にとろけるスライス,"スライスチーズ(4枚（スイスのグリュイエルチーズがおすすめ）),チキンブイヨン(1片),白米(1.5カップくらい？)"
If you want to change the prompts used during training, please modify the formatting_func_.+ function in data_preprocessing.py. The following function is a sample for Cookpad.
def formatting_func_cookpad(example):
output_texts = [f"# ユーザ\n## タイトル\n{example['title'][i]}\n\n# アシスタント\n## 食材\n{example['ingredients'][i]}\n## 作り方\n{example['steps'][i]}" for i in range(len(example))]
return output_textsAn example of a dataset with the formatting_func_cookpad function applied is shown below.
# ユーザ
豚の角煮
# アシスタント
## 食材
しょうが（お好みで）、ニンニク（お好みで）、ねぎ（１本）、豚肉（バラのブロック２パック）、砂糖（小さじ１から２くらい）、酒（たくさん（安い日本酒でいい））、醤油（適量（味見しながらね））、みりん（大さじ３くらい）
## 作り方
鍋に、水とたっぷりのお酒、ねぎの使わない葉の部分、しょうがの皮、にんにくを入れて、２，３時間煮込みます。その間、あくや浮いてきた脂を丁寧に取りましょう。煮込んだお肉を、いったん水で洗いましょう。落とし蓋をして１時間。食べるちょっと前にねぎを入れて、味がついたらたべましょう。写真のは、ちんげん菜を入れてみました。鍋に、豚肉をいれて、酒、砂糖、みりん、醤油、しょうが（薄切り）、にんにくで煮込みます。
Warning
Some tokenizers (e.g., LLaMA 2) tokenize sequences in ways that differ from the usual methods. As a result, the provided code may not work correctly for training.
For more details, please refer to this website.
Tip
To resolve this issue, please use the following code.
from trl import DataCollatorForCompletionOnlyLM
data_collator = DataCollatorForCompletionOnlyLM(
response_template=tokenizer.encode("\n# アシスタント\n", add_special_tokens=False)[2:],
instruction_template=tokenizer.encode("# ユーザ\n", add_special_tokens=False),
mlm=False,
tokenizer=tokenizer
)This program operates using the Causal Language Model (CLM) available from Hugging Face. The CLM is widely used for text generation.
| Major Category | Subcategory | Sub-subcategory | Paper | Usage |
|---|---|---|---|---|
| Quantization | 8 bit | python run/cookpad.py --load-in-8bit |
||
| Quantization | 4 bit | python run/cookpad.py --load-in-4bit |
||
| Flash Attention | Flash Attention 2 | FlashAttention-2: Faster Attention with Better Parallelism and Work Partitioning | python run/cookpad.py --attn-implementation flash_attention_2 --torch-dtype float16 or python run/cookpad.py --attn-implementation flash_attention_2 --torch-dtype bfloat16 |
|
| PEFT | Soft prompts | Prompt Tuning | The Power of Scale for Parameter-Efficient Prompt Tuning | python run/cookpad.py --peft-type PROMPT_TUNING --prompt-tuning-init TEXT --prompt-tuning-init-text 料理のタイトルから料理の材料と手順を予測する。 |
| PEFT | Soft prompts | P-Tuning | GPT Understands, Too | python run/cookpad.py --peft-type P_TUNING --encoder-hidden-size 768 |
| PEFT | Soft prompts | Prefix Tuning | Prefix-Tuning: Optimizing Continuous Prompts for Generation | python run/cookpad.py --peft-type PREFIX_TUNING --encoder-hidden-size 768 |
| PEFT | Adapters | LoRA | LoRA: Low-Rank Adaptation of Large Language Models | python run/cookpad.py --peft-type LORA --target-modules all-linear |
| PEFT | Adapters | AdaLoRA | Adaptive Budget Allocation for Parameter-Efficient Fine-Tuning | python run/cookpad.py --peft-type ADALORA |
| PEFT | Adapters | BOFT | Parameter-Efficient Orthogonal Finetuning via Butterfly Factorization | python run/cookpad.py --peft-type BOFT --target-modules all-linear |
| PEFT | Adapters | Llama-Adapter | LLaMA-Adapter: Efficient Fine-tuning of Language Models with Zero-init Attention | python run/cookpad.py --peft-type ADAPTION_PROMPT |
| PEFT | IA3 | Few-Shot Parameter-Efficient Fine-Tuning is Better and Cheaper than In-Context Learning | python run/cookpad.py --peft-type IA3 --target-modules all-linear --feedforward-modules all-linear |
|
| PEFT | Adapters | LoHa | FedPara: Low-Rank Hadamard Product for Communication-Efficient Federated Learning | python run/cookpad.py --peft-type LOHA --target-modules all-linear |
| PEFT | Adapters | LoKr | Navigating Text-To-Image Customization:From LyCORIS Fine-Tuning to Model Evaluation | python run/cookpad.py --target-modules all-linear |
| PEFT | Adapters | OFT | Controlling Text-to-Image Diffusion by Orthogonal Finetuning | python run/cookpad.py --peft-type OFT --target-modules all-linear |
| PEFT | Polytropon | Combining Modular Skills in Multitask Learning | python run/cookpad.py --peft-type POLY --target-modules all-linear |
|
| PEFT | Layernorm Tuning | Tuning LayerNorm in Attention: Towards Efficient Multi-Modal LLM Finetuning | python run/cookpad.py --peft-type LN_TUNING --target-modules all-linear |
|
| PEFT | FourierFT | Parameter-Efficient Fine-Tuning with Discrete Fourier Transform | python run/cookpad.py --peft-type FOURIERFT --target-modules all-linear |
|
| Generation Strategy | Greedy Decoding | python run/cookpad.py |
||
| Generation Strategy | Multinomial Sampling | python run/cookpad.py --do-sample |
||
| Generation Strategy | Beam-Search Decoding | python run/cookpad.py --num-beams 2 |
||
| Generation Strategy | Beam-Search Multinomial Sampling | python run/cookpad.py --do-sample --num-beams 2 |
||
| Generation Strategy | Contrastive Search | A Contrastive Framework for Neural Text Generation | python run/cookpad.py --penalty-alpha 0.5 |
|
| Generation Strategy | Diverse Beam-Search Decoding | Diverse Beam Search: Decoding Diverse Solutions from Neural Sequence Models | python run/cookpad.py --num-beams 2 --num-beam-groups 2 |
|
| Generation Strategy | Assisted Decoding | python run/cookpad.py --prompt-lookup-num-tokens 2 |
||
| Generation Strategy | DoLa Decoding | DoLa: Decoding by Contrasting Layers Improves Factuality in Large Language Models | python run/cookpad.py --dola-layers low |
bash main.shTo perform inference using the fine-tuned model, execute the following code. Replace checkpoint in the code with your path, and title with the title of the dish you want to generate.
import torch
from transformers import AutoTokenizer, AutoModelForCausalLM, GenerationConfig
checkpoint = "YOUR_PATH/checkpoint-X"
tokenizer = AutoTokenizer.from_pretrained(checkpoint)
model = AutoModelForCausalLM.from_pretrained(checkpoint, device_map="auto")
generation_config = GenerationConfig()
with torch.no_grad():
title = "料理のタイトル"
input_text = f"# ユーザ\n{title}\n\n# アシスタント\n"
input_text = tokenizer(input_text, add_special_tokens=True, return_tensors="pt").to(model.device)
output_text = model.generate(**input_text, generation_config=generation_config)
output_list = [tokenizer.decode(output_text[i], skip_special_tokens=True) for i in range(len(output_text))]