Skip to content

Latest commit

 

History

History
58 lines (43 loc) · 2.18 KB

README.md

File metadata and controls

58 lines (43 loc) · 2.18 KB

Kwansi Example Implementation

This repository contains an example implementation of Kwansi, a wrapper for DSPy that makes its optimizers easier to use. You can find Kwansi here: https://github.com/baloise/kwansi, which also includes a documentation about its core functions.

Overview

This example demonstrates how to use Kwansi to optimize a tweet creation task. It covers:

  • Setting up the environment
  • Preparing training data
  • Creating a task
  • Defining metrics
  • Running the optimizer
  • Testing the optimized model

Installation

  1. Clone this repository:

    git clone https://github.com/lordamp/kwansi-example.git
cd kwansi-example

  2. Install the required packages:

    pip install -r requirements.txt

  3. Set up your OpenAI API key in a .env file:

    OPENAI_API_KEY=your-api-key-here

Usage

  1. Run the main script:

    python example_implementation.py

  2. For a step-by-step walkthrough, see the example_implementation.ipynb Jupyter notebook.

File Structure

  • example_implementation.py: The example implementation of Kwansi. Run this to execute the whole example.
  • example_implementation.ipynb: The step-by-step walkthrough of the example implementation as a Jupyter notebook.
  • components/: Custom task and evaluator definitions
    • task.py: The task instructions we're trying to optimize
    • assessors.py: The assessor evaluators used to evaluate the task (LLM-based evaluators)
    • metrics.py: The metrics used to evaluate the task (binary metrics)
    • custom_combiner.py: A custom combiner function for the evaluators
  • data/: Example data for the task
    • example_data.json: The prepared data used for training the optimizer
  • outputs/: Directory for storing the compiled instructions (empty at the start)

Further Documentation

For more detailed information about Kwansi, please visit the Kwansi documentation. If you want to learn more about the underlying DSPy framework, please visit the DSPy documentation.