Keep this in mind when you do Model distillation! New paper from @GoogleDeepMind confirms that teacher hacking in model distillation with offline datasets! Offline Distillation refers to generating synthetic data (logits) from offline, fixed prompts with a teach model and then training a student model on it. Teacher hacking is when a student optimizes for mimicking, imitate the teacher's imperfect behavior, rather than learning the true underlying task, reducing its generalization on unseen tasks. Solutions: 1️⃣ Online (Recommended): In Online or on-policy Knowledge Distillation, a student learns from a teacher during training by minimizing the distribution of samples dynamically generated by the student with a forward KL Term between the teacher and student. 2️⃣ Diverse Offline (If Online is Infeasible): Create a diverse offline dataset. Prioritize a wide variety of prompts. If prompt variety is limited, generate multiple responses per prompt from the Teacher. Avoid small, static datasets with single responses. Insights 🔍 Teacher hacking emerges when using fixed offline datasets 🌐 Online data generation effectively prevents teacher hacking by maintaining response diversity 🎯 Higher prompt diversity reduces teacher hacking more than multiple responses per prompt ⏰ Limiting training epochs helps avoid teacher hacking with offline data 📈 Using online data generation (Teacher or Student responses generated during training) prevents teacher hacking. 🔄 Multiple completions per prompt (2x-3x) bridges the gap between offline and online performance 🔎 Teacher hacking can be detected by monitoring proxy metrics (Student-Teacher distance).