Prompt engineering на собеседовании Data Scientist

Зачем prompt engineering

Right prompt → 2-10× quality на same model. Cost-effective alternative fine-tuning.

System / user / assistant

Standard message structure.

system: "You are a helpful assistant что translates English to Russian."
user: "Hello, how are you?"
assistant: "Привет, как дела?"

System message — sets behavior, persona, constraints.

Few-shot

Include examples в prompt.

Translate to formal Russian:

"Hi" → "Здравствуйте"
"Bye" → "До свидания"
"Thanks" → "Спасибо"

"How are you?" → ?

In-context learning. 1-5 examples often suffice.

Chain-of-thought

«Think step by step».

Question: A train leaves NYC at 100 mph. ...

Answer: Let me think step by step.
Step 1: ...
Step 2: ...
Therefore: ...

Improves reasoning accuracy 20-30% на math / logic.

Zero-shot CoT. Просто add «Let's think step by step» — works.

Self-consistency. Generate multiple CoT solutions, vote.

Structured output

Return JSON / Pydantic schema.

Return JSON {"name": str, "age": int, "interests": [str]}.

Input: "John is 30 years old and likes hiking and reading."

Strict mode в OpenAI / Anthropic API — гарантирует JSON valid.

Function calling. API-level structure.

Связанные темы

• BERT vs GPT для DS
• AI agents для DS
• Few-shot learning для DS
• RAG на собесе DS
• Подготовка к собесу Data Scientist

FAQ

Это официальная информация?

Нет. Статья основана на работах Wei 2022 (CoT), Yao 2023 (ReAct), документации OpenAI / Anthropic.

Тренируйте Data Science — откройте тренажёр с 1500+ вопросами для собесов.