๐ TwinWeaver Hackathon Challenges¶
Welcome to the TwinWeaver Hackathon! These interactive notebooks are designed to test your understanding of clinical data processing and LLM fine-tuning for medical forecasting.
๐ฏ Challenge Format¶
Unlike the standard tutorials, these notebooks require active participation:
- ๐ด TODO Sections: Code cells marked with
# TODO:require you to write code - โ Quiz Questions: Test your understanding before moving forward
- ๐งช Experiments: Design and run your own experiments
- ๐ Checkpoints: Verify your solutions before proceeding
- ๐ Bonus Challenges: Optional advanced tasks for extra credit
๐ Available Challenges¶
Challenge 1: Data Preparation (01_data_preparation_challenge.ipynb)¶
Difficulty: โญโญ (Intermediate) Time: 45-60 minutes
Learn to transform raw clinical data into instruction-tuning format by: - Configuring the data pipeline from scratch - Understanding split strategies for patient timelines - Debugging common configuration errors - Experimenting with different forecasting targets
Challenge 2: End-to-End LLM Fine-tuning (02_llm_finetuning_challenge.ipynb)¶
Difficulty: โญโญโญ (Advanced) Time: 90-120 minutes
Master the complete LLM fine-tuning workflow by: - Setting up training data generation - Configuring LoRA hyperparameters - Analyzing training dynamics - Running inference experiments - Evaluating model predictions
๐ Getting Started¶
- Make sure you have TwinWeaver installed:
pip install twinweaver - For Challenge 2, also install:
pip install twinweaver[fine-tuning-example] - Open the challenge notebook and follow the instructions
- Don't peek at the solutions until you've tried!
๐ก Tips for Success¶
- Read the context: Each section provides background information
- Check the hints: Stuck? Look for the ๐ก hint markers
- Run checkpoints: Use the validation cells to verify your code
- Experiment freely: The bonus sections encourage exploration
๐ Scoring (Optional)¶
If running as a competition: - Each TODO section: 10 points - Each quiz question: 5 points - Bonus challenges: 15-25 points each - Clean, well-documented code: up to 10 bonus points
Good luck! ๐