Hack your way into LLMs

Dive deep into the intriguing world of mechanistic interpretability at EPFL! 🚀 With SAIL hosting an engaging hackathon, you'll have the chance to explore a curated adversarial dataset with five innovative coding models. Your mission? Understand what's happening internally when these models face potentially harmful requests.

Two Tracks to Explore:

🔍 Probes as Monitors: Classify internal activations to detect harmful compliance before the model generates potentially risky outputs.
🎯 Attribution: Identify which specific inputs push a model toward supporting an attacker.

What’s Included:

• Claude Code Credits
• GPU Instances
• Mentorship
• Food throughout the weekend

This is your chance to achieve a publishable outcome in just 48 hours! 🏆

Prizes Awaiting:

💰 CHF 1,000
💰 CHF 500
💰 CHF 300

Bring your PyTorch fluency and curious mind focused on model internals — that's all you need to participate!

Important:

📝 Register by May 6th
📅 Event Date: May 9-10
👥 Team Size: Up to 5 members

👉 Register Here

Join us for a thrilling weekend of innovation and discovery!