ClashAI-300k

ClashAI-300k

ClashAI-300k is a lightweight Transformer language model with approximately 300,000 parameters, designed to answer questions about Clash Royale in Portuguese.

This project was created as an educational experiment to demonstrate how a small language model can be trained from scratch using PyTorch, a custom tokenizer, and a handcrafted dataset.

Features

• 🧠 Approximately 300K parameters
• 🇧🇷 Portuguese language
• 🎮 Specialized in Clash Royale
• ⚡ Lightweight and fast inference
• 🔤 Custom BPE tokenizer
• 🔥 Built entirely with PyTorch
• 📚 Trained from scratch (no pretrained weights)

Intended Use

ClashAI-300k is intended for:

• Answering Clash Royale questions
• Educational purposes
• Learning how Transformer models work
• Small AI experiments
• Running on low-end hardware

It is not intended to replace large language models or provide general-purpose knowledge.

Model Details

Training

The model was trained from scratch using a manually curated dataset containing hundreds of Clash Royale question-answer pairs.

Training includes:

• Custom tokenizer training
• Sequence-to-sequence language modeling
• Cross-entropy loss optimization
• Autoregressive text generation

Example

Input

Como conseguir ouro?

Output

Você pode conseguir ouro vencendo batalhas, abrindo baús, participando de eventos e doando cartas ao clã.

---

Input

Como derrotar um Corredor?

Output

Use construções defensivas para desviar o Corredor e tropas de alto dano para eliminá-lo rapidamente.

Limitations

ClashAI-300k is intentionally small.

Some limitations include:

• Limited world knowledge
• Specialized only in Clash Royale
• May generate incorrect or repetitive responses
• Performs best on questions similar to its training data

Future Improvements

Planned improvements include:

• Larger training dataset
• Better sampling strategies
• Larger context window
• Improved architecture
• More gameplay knowledge
• Better generalization

Disclaimer

This is an independent educational project.

Clash Royale is a trademark of Supercell. This project is not affiliated with, endorsed by, or sponsored by Supercell.

License

MIT License