You already know: SFT data quality (Supervised fine-tuning §5); preference pairs and DPO (Preference optimization); verifiers, pass@k and reward hacking (RL for LLMs); near-duplicate detection and benchmark contamination (LLM pretraining §1); softmax temperature (D2 Information theory).
Why this matters for Atlas. Most post-training data in modern models is synthetic: generated by models, then filtered. It is how small models inherit the skills of large ones (distillation), how instruction datasets reach millions of examples (Self-Instruct), and how reasoning models get worked solutions (rejection sampling with verifiers). It is also full of traps: generators repeat themselves, filters skew the data toward easy examples, models trained on their own outputs slowly forget the real world, and synthetic data can quietly leak your benchmark. This lesson covers the methods together with the failure modes, each one measured.
1. Teacher–student distillation
The question. We have a strong but expensive model. Can a small, cheap model learn to do what it does?
Start from nothing. In distillation, a large teacher labels data and a small student trains on those labels. Two flavours: