Using this dataset for LLM fine-tuning
Hi everyone, I am trying to fine-tune a Tiny LLAMA on a custom dataset and I came across this dataset. My idea is to fine-tune the base model to then use it as a chat model focused on recipe generation based on the prompt given which includes a list of ingredients that the user has.
What is the best way to structure dataset in the necessary format? I am looking for something along the lines of:
"""Format the prompt using the <|user|> and <|assistant|> format“""
<|user|>\n3. Heat the vegetables in boiling water for the recommended time.\n<|assistant|>\nI do not have information about the specific type of vegetables being referred to. However, here are general instructions for boiling most vegetables:\n\n1. Wash the vegetables thoroughly with clean water.\n2. Cut the vegetables into small or medium-sized pieces.\n3. Bring a pot of water to boil on the stove.\n4. Add a pinch of salt to the boiling water.\n5. Add the vegetables to the boiling water.\n6. Cook the vegetables for the recommended time (see cooking instructions on the package or look up cooking times for specific vegetables online).\n7. Test the vegetables for doneness using a fork or a knife. They should be tender but not overcooked and mushy.\n8. Once the vegetables are cooked to your desired tenderness, remove them from the boiling water using a slotted spoon or a strainer.\n9. Drain the vegetables and serve hot with your favorite seasonings or sauce.
I was thinking of using a template tokenizer to then apply_chat_template, but I wanted to ask the community if this is the right procedure or I should be doing this manually