dadukhankevin / llgenetics

Large Language Genetics

Large Language Genetics (LLGenetics)

LLGenetics is a Python library that combines large language models (LLMs) with genetic algorithms to evolve and optimize text prompts. It provides a set of customizable layers and components to build genetic algorithms specifically tailored for working with text data and LLMs.

Features

• Integration with the Anthropic API for using the Claude model
• Customizable gene pools for generating initial text populations
• Mutation layers for introducing variations in the text
• Crossover layers for combining and generating new text offspring
• Fitness evaluation and selection layers
• Support for both LLM-based and pure text-based genetic operations

Installation

To install LLGenetics, clone the repository from GitHub:

git clone https://github.com/dadukhankevin/LLGenetics.git

Usage

Here's a basic example of how to use LLGenetics:

from llgenetics import Claude, LLMGenePool, PopulateLLM, MutationLayer, CrossoverLayer, SortByFitnessLayer, CapPopulationLayer, Sequential

# Initialize the LLM (Claude) with API key and prompt templates
llm = Claude(api_key="your_api_key", system_prompt=DEFAULT_MUTATION_SYSTEM_PROMPT, template=MUTATION_PROMPT)

# Create a gene pool for generating initial text population
gene_pool = LLMGenePool(llm=llm, populate_prompt=POPULATE_PROMPT)

# Define the layers of the genetic algorithm
layers = [
    PopulateLLM(gene_pool=gene_pool, population_cap=10),
    MutationLayer(llm=llm, individual_selection=lambda individuals: individuals[:5]),
    CrossoverLayer(llm=llm, individual_selection=lambda individuals: individuals[:5]),
    SortByFitnessLayer(),
    CapPopulationLayer(max_population=10)
]

# Create the genetic algorithm environment
environment = Sequential(layers=layers, fitness_function=your_fitness_function)

# Run the genetic algorithm for a specified number of generations
environment.run(generations=10)

Customization

LLGenetics provides various customization options:

• Implement your own fitness function to evaluate the quality of the generated text prompts
• Adjust the mutation and crossover parameters to control the level of variation introduced
• Define custom selection strategies for choosing individuals for mutation and crossover
• Experiment with different prompt templates to optimize the text generation process
• Right now Claude Haiku is the only LLM fast and smart enough to be useful in GAs, but more will be added.

Contributing

Contributions to LLGenetics are welcome! If you find any bugs, have suggestions for improvements, or want to add new features, please open an issue or submit a pull request on the GitHub repository.

License

LLGenetics is open-source and released under the MIT License.