Model card на собеседовании Data Scientist

Что такое model card

Documentation для ML model. Like dataset readme, но для model.

Mitchell 2019 (Google). Standard в Hugging Face Hub.

Структура

Template:

# Model Card: model_name

## Model Details
- Architecture, version, authors.

## Intended Use
- Primary uses.
- Out-of-scope uses.

## Training Data
- Source, size, composition.

## Evaluation
- Metrics, benchmarks.
- Per-subgroup performance.

## Ethical Considerations
- Biases.
- Risks.
- Mitigations.

## Caveats
- Limitations.
- Recommendations.

Intended use

Critical — declare what model для.

Primary uses. «Email spam classifier для В2С email».

Out of scope. «Не использовать для legal advice classification».

Helps users assess fit к their use case. Prevents misuse.

Performance metrics

Не just average. Disaggregated:

• Per language.
• Per demographic group.
• Per data source.

| Group | Accuracy |
| RU users | 92% |
| Other | 87% |

Reveals biases, areas needing improvement.

Ethical considerations

Bias sources. Training data demographics, label biases.

Privacy. PII в training? PII в outputs (training data leakage)?

Misuse potential. Could used для discrimination, manipulation?

Mitigations. Что мы сделали — fair sampling, bias testing, content filters.

В regulated industries (finance, medicine) — model card is mandatory artifact.

Связанные темы

• Bias и fairness для DS
• MLflow и DVC для DS
• Hallucinations и LLM evals для DS
• SHAP и interpretability для DS
• Подготовка к собесу Data Scientist

FAQ

Это официальная информация?

Нет. Статья основана на работе Mitchell 2019 «Model Cards for Model Reporting».

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