Hallucination Detection
ПродвинутыйtestingМинимальный контекст: 32K
Designs checks that catch unsupported or fabricated claims in LLM output before users see them. Covers grounding verification against retrieved context, claim decomposition and per-claim entailment checking, citation validation, self-consistency sampling, and abstention thresholds, plus how to measure detector precision and recall on a labeled set.
Варианты использования
- Verifying that RAG answers are grounded in retrieved passages
- Decomposing an answer into claims and checking each for support
- Validating that citations actually contain the cited statement
- Setting abstention thresholds so the model says "I don't know"
Пример промпта
Design a hallucination detection layer for our RAG assistant over internal documentation. Cover: 1. How to decompose an answer into individually checkable claims 2. How to score each claim for support against the retrieved chunks 3. How to validate citations point at text that actually supports the claim 4. Thresholds for rewriting versus abstaining 5. How I measure detector precision and recall, and what labeled data I need Include the implementation and note the latency and cost this adds per request.
Рекомендуемые модели
Совместимые инструменты
claude-codecursorkiroany
Модальности
Вход: text
→Выход: text, code
Похожие Skills
Автор
OpenModels Community