Numenigma: Crafting Math Riddles with TinyLlama 🤖💡

Welcome to Numenigma — a fun and creative project where we unlock the mysteries of numbers by generating intriguing math riddles using the lightweight TinyLlama model! In this blog post, I’ll take you through the journey of fine-tuning a language model to generate unique math riddles, explain why I chose TinyLlama, how the model works, and share useful links and hashtags to get involved. Let’s dive in!

🚀 Project Overview

Numenigma is designed to automatically generate math riddles that not only challenge your mind but also entertain you with clever arithmetic puzzles. The project consists of three main steps:

1. Dataset Preparation: We created a custom dataset of 50 unique math riddles.
2. Model Fine-Tuning: Using PyTorch and the Hugging Face Transformers library, we fine-tuned TinyLlama on our dataset.
3. Interactive Testing: A friendly Streamlit interface lets users generate and test new math riddles on demand.

By combining modern NLP techniques with a playful twist on math puzzles, Numenigma bridges the gap between technology and creativity. 🎉

🔍 Why TinyLlama?

Lightweight & Efficient

TinyLlama is a compact language model that provides an excellent balance between performance and computational efficiency. This makes it perfect for:

• Low-resource environments: It runs smoothly on modest hardware without needing massive GPU resources.
• Fast iterations: Its smaller size allows for quick fine-tuning and testing, enabling rapid experimentation.

How It Works

TinyLlama, like other transformer-based models, is built on self-attention mechanisms. Here’s a brief overview of how it generates text:

• Input Embeddings: The model converts text inputs into dense vector representations.
• Self-Attention Layers: These layers allow the model to weigh the importance of different words and phrases in the input.
• Output Generation: Using techniques like beam search or sampling (we use nucleus sampling with parameters such as top_k and top_p), the model generates a continuation from a given prompt.

In our project, we fine-tune TinyLlama on a dataset of math riddles so that it learns the style, structure, and logical progression required to craft both the question and answer parts of each riddle. This specialized fine-tuning empowers the model to produce coherent and engaging riddles on demand.

🛠️ Fine-Tuning & Testing Details

Fine-Tuning Process

• Dataset: We prepared a CSV file containing 50 unique math riddles.
• Training: The model was fine-tuned over several epochs using PyTorch and Hugging Face’s Trainer. We optimized hyperparameters such as learning rate, weight decay, and batch size to achieve stable convergence.
• Randomization: To maximize the learning potential on our small dataset, we randomize the training samples each epoch.

Testing Interface with Streamlit

We built an interactive web interface using Streamlit that allows you to:

• Input a custom prompt (e.g., “Riddle:”).
• Adjust generation parameters like max tokens, top_k, and top_p.
• View the generated math riddle complete with its answer instantly.

Check out the code snippet in the project repository for details on how the interface is set up!

🔗 Links & Resources

• GitHub Repository:

GitHub - buzzgrewal/Numenigma

• Kaggle:

https://www.kaggle.com/models/buzzgrewal/numenigma

🎉 Conclusion

Numenigma is more than just a project — it’s an exciting fusion of creative thinking and cutting-edge AI. With TinyLlama at its core, this project demonstrates that even small models can deliver big ideas! Whether you’re a math enthusiast, an AI hobbyist, or simply love a good riddle, I hope this project inspires you to explore the endless possibilities at the intersection of technology and art.

Feel free to fork the project, experiment with the code, and share your own unique riddles. Let’s unlock the mysteries of numbers together! 🔢✨

#Numenigma #MathRiddles #TinyLlama #NLP #MachineLearning #DeepLearning #AI #Streamlit #MathPuzzles #TechArt

Happy riddle crafting, and see you in the next post!