This project combines the worlds of Artificial Intelligence and poetry by creating a haiku generator. Using Rob Gonsalves' Deep Haiku Llama 2 model as a foundation, we trained our dataset to generate haikus. The project involves a blend of technologies including Google Colab, Flask server, and a Streamlit app, creating an AI-powered tool that crafts haikus in the traditional 5-7-5 syllable structure, a mixture of technology and the linguistic arts.
Our dataset, sourced from Kaggle (Haiku Dataset 1 and Haiku Dataset 2), comprises a diverse collection of haikus. It contains a comprehensive collection of haikus, providing a rich foundation for training the AI model. This dataset was integral in training the model to understand and generate haikus adhering to the traditional structure and thematic elements.
The core algorithm of this project is based on the Deep Haiku Llama 2 model developed by Rob Gonsalves. This model has been specially fine-tuned to generate rhythmic prose, particularly haikus. The algorithm processes the dataset through several steps, including punctuation addition, keyword extraction, quality filtering, and syllable counting, ensuring the generation of high-quality haikus.
- LlamaCPP Model: Employed for haiku generation in Google Colab.
- Flask: Serves as the backend server, handling requests for haiku generation.
- ngrok: Provides a public URL to access the Flask app over the internet.
- Streamlit: Used to create a user-friendly interface for interacting with the Flask server, allowing users to input topics and receive haikus along with sentiment analysis and word cloud visualization.
- Google Colab: Provides a cloud-based environment for running the LlamaCPP model due to GPU constraints.
- LlamaCPP Model Initialization and Haiku Generation: Configured the LlamaCPP model in Google Colab for generating haikus on various topics.
- Flask Web Server: Developed a Flask app with an endpoint
/generate_haikuto produce haikus based on user inputs. - ngrok Configuration: Enabled internet access to the Flask app through a public URL.
- Streamlit Application: Created a user interface for users to request haikus, featuring sentiment analysis and word cloud visualization.
- Quality Assessment and Prompt Tuning: Continuously working on improving the quality of the haikus through prompt optimization.
- Streamlit App Enhancements: Enhancing the UI/UX of the Streamlit application and refining features like sentiment analysis and word cloud visualization.
- GPU Resource Constraints: Currently running the project on Google Colab due to limited GPU resources, affecting scalability and performance.
The complete code and documentation for this project are available at HaikuGenerator on GitHub.
Special thanks to Kaggle for providing the datasets that were instrumental in training our model. Also, a note of gratitude to Rob Gonsalves for his foundational work on the Deep Haiku project and his contributions to the field.