LlaMa Lunch

Integrated with Meta's Large Language Model, LlaMa2, LlaMa Lunch is an interactive AI-cook assistant designed to inspire you with recipe ideas for your upcoming culinary endeavours.

Just input the ingredients you wish to work with, and let LlaMa Lunch seamlessly generate a customized recipe tailored to your preferences!

In the interest of convenience, LlaMaLunch leverages the API endpoint provided by Hugging Face rather than relying on the training code located in the train_code/ directory. But the behaviour is replicated as shown in the Jupyter Notebook.

We have detailed our efforts and attempts in train_code/.

Product Demo

All the user needs to do is input some ingredients and LlaMaLunch will generate a recipe using only the ingredients provided as shown below:

Getting Started

Run the backend server

1. In the project directory, cd llama-be
2. Run FLASK_APP=app.py flask run to start the backend server on localhost.

Run the frontend server

1. In the project directory, cd llama-fe
2. Run npm run start to start the server.
3. Open http://localhost:3000 to view it in the browser.

Model Parameters

The LlaMa2 model from Huggingface accepts 7 parameters: Message, System prompt, Max new tokens, Temperature, Top-p (nucleus sampling), Top-k, and Repitition penality.

• Message: The user's input prompt

• System prompt: A block of text appended before the Message to steer LlaMa2's responses. This helps to guide the model's subsequent behaviour through preemptive framing of the discussion paramters.

  In the context of LlaMa Lunch, we want to generate chef-like recipes that are feasible. As such, we include a system prompt of 'You are a chef brainstorming recipes'. This directive ensures that the generated recipes are not just creative but also within the realm of feasibility.

The decision to incorporate a persona (Chef) into the model was influenced by the insights from the LlaMa2 paper that details the type of prompts the model can handle:

• Max new tokens: Maximum number of tokens to generate. This determines the maximum length of the model's output. Setting a lower value limits the maximum number of tokens generated to avoid excessively long responses.

  To generate more elaborate and extended recipes, we have configured this value to its maximum setting.

• Temperature: Determines the randomness of the outputs. A higher temperature value generates a vastly different output. A low temperature value generates a more consistent and conservative responses.

  For LlaMa Lunch, our goal is to produce captivating recipes, so we opted for a higher temperature setting to enhance the diversity and creativity of our generated output.

• Repetition penalty: Penalizes or reduces the probability of generating tokens that have recently appeared in the generated text. A value greater than 1 discourages word repetition in the model's output.

  Similar to the rationale behind the value set for Temperature, we want to ensure creativity in recipes generated. Hence, this means reducing repetition by increasing the repetition penalty value.

• Top-p: Constraints the next token selection to the top-k most likely tokens at each step.

• Top-k: Limits the selection of tokens to a subset of the vocabulary with cumulative probability mass up to a threshold value. This helps in controlling the diversity of generated output.