Викторина, связанная с руководством · Средний Уровень
IFEval и тест на соответствие инструкциям
Поймите проверяемые ограничения IFEval, область оценки, воспроизводимость и ограничения как меру более широкого качества модели.
Вопрос 1 из 8
What capability does IFEval primarily measure?
Предварительный просмотр каждого вопроса8 вопросы
Ответы и пояснения появляются по мере прохождения теста.
- What capability does IFEval primarily measure?
- Factual correctness across all topics
- Compliance with specified objectively checkable instructions
- People’s writing-style preference
- External tool use
- Which is an example of an IFEval-style instruction?
- Mention a specified keyword at least three times
- Give the most insightful answer
- Be more thoughtful than another model
- Show empathy naturally
- What did the original IFEval paper report about its construction?
- 25,000 preference pairs
- One prompt per profession
- Only mathematical reasoning
- 25 types of verifiable instructions and around 500 prompts
- Why can a deterministic checker help with IFEval-style constraints?
- It can check an explicit rule like a count without a generative judge
- It proves the answer factual
- It makes every instruction unambiguous
- It removes the need to inspect test design
- What does an IFEval pass establish?
- The answer is complete for every user
- The model generalizes to all unseen instructions
- The response met the tested constraint or constraints
- The answer is factually correct
- Why preserve the official benchmark version and evaluator?
- Changes can alter what is measured and the score
- The evaluator controls model weights
- Version determines factual truth
- All teams must use one vendor
- If an answer meets a word-count rule but contains a false explanation, how should IFEval reflect that?
- The falsehood changes the word count
- The constraint can pass while factual quality is evaluated separately
- IFEval proves it correct
- It receives full marks for all capabilities
- How should a team use an IFEval score in a product decision?
- As the complete substitute for application testing
- Rank without inspecting failures
- Treat a high score as a safety certificate
- Use it as one scoped measure, then test application-specific needs