ByteDance Douyin E-commerce AI Infra

Training optimization

1. In the SFT stage, what model, data sources, data processing, and training framework did you use?
2. Walk through PPO training in detail: data prep, reward-model training, loss design for each model, and compute resources used.
3. What is the DPO training flow? How do PPO and DPO differ? Did you run comparison experiments?
4. How do you evaluate quality gains after fine-tuning?
5. Why might PPO outperform DPO? How do you interpret gains from multi-round DPO?
6. What is the principle of GRPO?
7. Why do you still need RLHF after SFT?

Fundamentals

8. How is CLIP trained?
9. Introduce several common deep-learning optimizers.
10. How does Multi-Head Attention work?
11. How does LoRA work? How does rank r affect training?
12. How is ViT trained?
13. What is the core design idea of Swin Transformer?
14. How do fast thinking and slow thinking work in Qwen3?

Project experience

15. Introduce project innovations and code details, including ablation experiment design.

Coding

16. Coding assessment.