Adaptation for Deep Learning Engineers
Adaptation is how a pretrained model is specialised for a task or behaviour. Supervised fine-tuning trains on example instructions and answers, parameter-efficient methods such as LoRA update a small number of weights, preference optimisation like DPO aligns outputs with human choices, and distillation transfers skill to a smaller model.
Stage: Language and Multimodal Models · Part of the Deep Learning Engineer roadmap · Last reviewed 19 November 2026
What to learn in Adaptation
- Prepare instruction data in a consistent chat template
- Fine-tune with LoRA and choose rank and target modules
- Explain how DPO uses preference pairs without a reward model
- Distil a large teacher model into a smaller student
- Generate and filter synthetic training data for a task
- Detect catastrophic forgetting after fine-tuning
Adaptation interview questions
How does LoRA reduce the cost of fine-tuning?
LoRA freezes the original weights and learns a low-rank update, the product of two small matrices, for selected layers. Only these few parameters and their optimiser state need training memory, and the update can be merged into the weights afterwards.
What is the difference between supervised fine-tuning and preference optimisation?
Supervised fine-tuning imitates given target answers. Preference optimisation uses pairs of preferred and rejected answers to shift the model toward better responses, shaping style, helpfulness or safety beyond what imitation data alone provides.
Where Adaptation fits in the Deep Learning Engineer roadmap
Adaptation belongs to the Language and Multimodal Models stage (advanced level), which takes about 3 hours. Text, speech and beyond.
Other units in this stage
Back to the full Deep Learning Engineer roadmap