Hallucination detection in Prompt Engineering / GenAI - Practice Problems & Coding Challenges

Challenge - 5 Problems

🧠 Conceptual

intermediate

What is hallucination in AI-generated text?

In the context of AI language models, what does the term 'hallucination' mean?

A The model repeats the same sentence multiple times in the output.

B The model produces text that is grammatically incorrect but factually accurate.

C The model generates text that is factually incorrect or unsupported by data.

D The model refuses to generate any output due to lack of data.

💡 Hint

Think about when AI makes up information that isn't true.

Hallucination in AI means the model creates information that looks plausible but is actually false or unsupported.

❓ Predict Output

intermediate

Detecting hallucination with confidence scores

Given a model output with confidence scores for each token, which output indicates a higher chance of hallucination?

Prompt Engineering / GenAI

tokens = ['The', 'capital', 'of', 'France', 'is', 'Berlin']
confidences = [0.99, 0.98, 0.97, 0.96, 0.95, 0.40]

# Which token likely indicates hallucination?

A The token 'The' with confidence 0.99

B The token 'Berlin' with confidence 0.40

C The token 'capital' with confidence 0.98

D The token 'France' with confidence 0.96

💡 Hint

Lower confidence tokens may indicate hallucination.

Lower confidence on a token, especially when it contradicts known facts, suggests hallucination.

❓ Model Choice

advanced

Choosing a model architecture to reduce hallucination

Which model architecture is best suited to reduce hallucination in generated text by grounding outputs on external knowledge?

A A retrieval-augmented generation model that uses external databases during generation

B A convolutional neural network trained on images

C A standard transformer language model trained only on text data

D A recurrent neural network trained on random noise

💡 Hint

Models that check facts during generation help reduce hallucination.

Retrieval-augmented models access external knowledge to verify facts, reducing hallucination.

❓ Metrics

advanced

Evaluating hallucination with automatic metrics

Which automatic metric is most appropriate to measure hallucination in AI-generated summaries?

A FactCC score measuring factual consistency between summary and source document

B BLEU score comparing generated text to reference text

C ROUGE score comparing generated summary to reference summary

D Perplexity measuring how well the model predicts the next word

💡 Hint

Look for metrics that check factual correctness, not just similarity.

FactCC is designed to detect factual inconsistencies, making it suitable for hallucination evaluation.

🔧 Debug

expert

Identifying hallucination cause in model output

You have a language model that often hallucinates facts in its answers. Which debugging step is most likely to help reduce hallucination?

A Remove all stop words from the training data

B Increase the model size without changing training data

C Reduce the learning rate to zero during training

D Fine-tune the model on a high-quality, fact-checked dataset

💡 Hint

Improving training data quality helps reduce hallucination.

Fine-tuning on fact-checked data helps the model learn accurate information and reduces hallucination.